llama.cpp/examples/train-text-from-scratch/train-text-from-scratch.cpp

2272 lines
97 KiB
C++
Raw Normal View History

train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#include "ggml.h"
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
#include "ggml-alloc.h"
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
#include "common.h"
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#include "llama.h"
#include <unordered_map>
#include <vector>
#include <cassert>
#include <climits>
#include <cstring>
#include <cstdarg>
#include <ctime>
#include <random>
#include <stdexcept>
#include <algorithm>
#include <string>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
struct random_normal_distribution {
std::mt19937 gen;
std::normal_distribution<float> rd;
float min;
float max;
};
struct random_uniform_distribution {
std::mt19937 gen;
std::uniform_real_distribution<float> rd;
};
void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
rnd->gen = std::mt19937(seed);
rnd->rd = std::normal_distribution<float>{mean, std};
rnd->min = min;
rnd->max = max;
}
void init_random_uniform_distribution(struct random_uniform_distribution * rnd, int seed, float min, float max) {
rnd->gen = std::mt19937(seed);
rnd->rd = std::uniform_real_distribution<float>{min, max};
}
int clamp(const int v, const int min, const int max) {
return ((v < min) ? (min) : (v > max) ? (max) : v);
}
float fclamp(const float v, const float min, const float max) {
return ((v < min) ? (min) : (v > max) ? (max) : v);
}
float frand() {
return (float)rand()/(float)RAND_MAX;
}
float frand_normal(struct random_normal_distribution * rnd) {
return fclamp(rnd->rd(rnd->gen), rnd->min, rnd->max);
}
float frand_uniform(struct random_uniform_distribution * rnd) {
return rnd->rd(rnd->gen);
}
struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
float scale = 1.0f; // xavier
switch (tensor->n_dims) {
case 1:
scale /= sqrtf(tensor->ne[0]);
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
*dst = scale * frand_normal(rnd);
}
break;
case 2:
scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
*dst = scale * frand_normal(rnd);
}
}
break;
case 3:
scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
*dst = scale * frand_normal(rnd);
}
}
}
break;
case 4:
scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
*dst = scale * frand_normal(rnd);
}
}
}
}
break;
default:
assert(false);
};
return tensor;
}
struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
switch (tensor->n_dims) {
case 1:
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
*dst = frand_uniform(rnd);
}
break;
case 2:
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
*dst = frand_uniform(rnd);
}
}
break;
case 3:
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
*dst = frand_uniform(rnd);
}
}
}
break;
case 4:
for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
*dst = frand_uniform(rnd);
}
}
}
}
break;
default:
assert(false);
};
return tensor;
}
struct my_llama_hparams {
uint32_t n_vocab = 32000;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
uint32_t n_ctx = 512;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
uint32_t n_embd = 4096;
uint32_t n_head = 32;
uint32_t n_layer = 32;
uint32_t n_rot = 64;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
uint32_t n_ff = 11008;
// float f_norm_eps = 1e-5; // falcon
float f_norm_rms_eps = 1e-5; // llama
float rope_freq_base = 10000.0f;
float rope_freq_scale = 1.0f;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
};
struct my_llama_layer {
// normalization
struct ggml_tensor * attention_norm;
// attention
struct ggml_tensor * wq;
struct ggml_tensor * wk;
struct ggml_tensor * wv;
struct ggml_tensor * wo;
// normalization
struct ggml_tensor * ffn_norm;
// ff
struct ggml_tensor * w1;
struct ggml_tensor * w2;
struct ggml_tensor * w3;
};
struct my_llama_model {
struct ggml_context * ctx = NULL;
my_llama_hparams hparams;
struct ggml_tensor * tok_embeddings;
struct ggml_tensor * norm;
struct ggml_tensor * output;
std::vector<my_llama_layer> layers;
uint32_t train_its = 0;
uint32_t train_samples = 0;
uint32_t train_tokens = 0;
};
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// gguf constants
const char * LLM_KV_OPTIMIZER_TYPE = "optimizer.type";
const char * LLM_KV_OPTIMIZER_TYPE_ADAM = "adam";
const char * LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs";
const char * LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version";
const char * LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count";
const char * LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count";
const char * LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count";
const char * LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized";
const char * LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss";
const char * LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss";
const char * LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count";
const char * LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count";
const char * LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss";
const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step";
const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j";
const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k";
const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end";
const char * LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count";
const char * LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments";
const char * LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments";
const char * LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s";
const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y";
const char * LLM_KV_TRAINING_FILE_VERSION = "training.file_version";
const char * LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count";
const char * LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count";
const char * LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count";
// gguf constants (sync with gguf.py)
const char * LLM_KV_GENERAL_ARCHITECTURE = "general.architecture";
const char * LLM_KV_GENERAL_FILE_TYPE = "general.file_type";
const char * LLM_KV_CONTEXT_LENGTH = "%s.context_length";
const char * LLM_KV_EMBEDDING_LENGTH = "%s.embedding_length";
const char * LLM_KV_BLOCK_COUNT = "%s.block_count";
const char * LLM_KV_FEED_FORWARD_LENGTH = "%s.feed_forward_length";
const char * LLM_KV_ATTENTION_HEAD_COUNT = "%s.attention.head_count";
const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon";
const char * LLM_KV_ROPE_DIMENSION_COUNT = "%s.rope.dimension_count";
const char * LLM_KV_ROPE_FREQ_BASE = "%s.rope.freq_base"; // TODO load in llama.cpp
const char * LLM_KV_ROPE_SCALE_LINEAR = "%s.rope.scale_linear";
const char * LLM_KV_TOKENIZER_MODEL = "tokenizer.ggml.model";
const char * LLM_KV_TOKENIZER_LIST = "tokenizer.ggml.tokens";
const char * LLM_KV_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type";
const char * LLM_KV_TOKENIZER_SCORES = "tokenizer.ggml.scores";
const char * LLM_KV_TOKENIZER_MERGES = "tokenizer.ggml.merges";
const char * LLM_KV_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id";
const char * LLM_KV_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id";
const char * LLM_KV_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id";
const char * LLM_KV_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id";
const char * LLM_KV_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id";
const char * LLM_TENSOR_TOKEN_EMBD = "token_embd";
const char * LLM_TENSOR_OUTPUT_NORM = "output_norm";
const char * LLM_TENSOR_OUTPUT = "output";
const char * LLM_TENSOR_ATTN_NORM = "blk.%d.attn_norm";
const char * LLM_TENSOR_ATTN_Q = "blk.%d.attn_q";
const char * LLM_TENSOR_ATTN_K = "blk.%d.attn_k";
const char * LLM_TENSOR_ATTN_V = "blk.%d.attn_v";
const char * LLM_TENSOR_ATTN_OUT = "blk.%d.attn_output";
const char * LLM_TENSOR_FFN_NORM = "blk.%d.ffn_norm";
const char * LLM_TENSOR_FFN_GATE = "blk.%d.ffn_gate";
const char * LLM_TENSOR_FFN_DOWN = "blk.%d.ffn_down";
const char * LLM_TENSOR_FFN_UP = "blk.%d.ffn_up";
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
void print_params(struct my_llama_hparams * params) {
printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
printf("%s: n_ctx: %d\n", __func__, params->n_ctx);
printf("%s: n_embd: %d\n", __func__, params->n_embd);
printf("%s: n_head: %d\n", __func__, params->n_head);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
printf("%s: n_ff: %d\n", __func__, params->n_ff);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
printf("%s: n_layer: %d\n", __func__, params->n_layer);
printf("%s: n_rot: %d\n", __func__, params->n_rot);
}
void init_model(struct my_llama_model * model) {
const auto & hparams = model->hparams;
const uint32_t n_embd = hparams.n_embd;
const uint32_t n_layer = hparams.n_layer;
const uint32_t n_vocab = hparams.n_vocab;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const uint32_t n_ff = hparams.n_ff;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
struct ggml_context * ctx = model->ctx;
model->train_its = 0;
model->train_samples = 0;
model->train_tokens = 0;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<char> tn_buf;
tn_buf.resize(GGML_MAX_NAME);
auto tn = [&tn_buf](const char * key) -> const char * {
snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
return tn_buf.data();
};
auto tni = [&tn_buf](const char * key, int bid) -> const char * {
snprintf(tn_buf.data(), tn_buf.size(), key, bid);
std::string s = tn_buf.data();
snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
return tn_buf.data();
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model->output = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_name(model->tok_embeddings, tn(LLM_TENSOR_TOKEN_EMBD));
ggml_set_name(model->norm, tn(LLM_TENSOR_OUTPUT_NORM));
ggml_set_name(model->output, tn(LLM_TENSOR_OUTPUT));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
model->layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
auto & layer = model->layers[i];
layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_name(layer.wq, tni(LLM_TENSOR_ATTN_Q, i));
ggml_set_name(layer.wk, tni(LLM_TENSOR_ATTN_K, i));
ggml_set_name(layer.wv, tni(LLM_TENSOR_ATTN_V, i));
ggml_set_name(layer.wo, tni(LLM_TENSOR_ATTN_OUT, i));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_name(layer.ffn_norm, tni(LLM_TENSOR_FFN_NORM, i));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_name(layer.w1, tni(LLM_TENSOR_FFN_GATE, i));
ggml_set_name(layer.w2, tni(LLM_TENSOR_FFN_DOWN, i));
ggml_set_name(layer.w3, tni(LLM_TENSOR_FFN_UP, i));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
void set_param_model(struct my_llama_model * model) {
const auto& hparams = model->hparams;
const uint32_t n_layer = hparams.n_layer;
struct ggml_context* ctx = model->ctx;
ggml_set_param(ctx, model->tok_embeddings);
ggml_set_param(ctx, model->norm);
ggml_set_param(ctx, model->output);
for (uint32_t i = 0; i < n_layer; ++i) {
auto & layer = model->layers[i];
ggml_set_param(ctx, layer.attention_norm);
ggml_set_param(ctx, layer.wq);
ggml_set_param(ctx, layer.wk);
ggml_set_param(ctx, layer.wv);
ggml_set_param(ctx, layer.wo);
ggml_set_param(ctx, layer.ffn_norm);
ggml_set_param(ctx, layer.w1);
ggml_set_param(ctx, layer.w2);
ggml_set_param(ctx, layer.w3);
}
}
void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
const auto & hparams = model->hparams;
const uint32_t n_layer = hparams.n_layer;
struct random_normal_distribution rnd;
init_random_normal_distribution(&rnd, seed, mean, std, min, max);
randomize_tensor_normal(model->tok_embeddings, &rnd);
randomize_tensor_normal(model->norm, &rnd);
randomize_tensor_normal(model->output, &rnd);
for (uint32_t i = 0; i < n_layer; ++i) {
auto & layer = model->layers[i];
randomize_tensor_normal(layer.attention_norm, &rnd);
randomize_tensor_normal(layer.wq, &rnd);
randomize_tensor_normal(layer.wk, &rnd);
randomize_tensor_normal(layer.wv, &rnd);
randomize_tensor_normal(layer.wo, &rnd);
randomize_tensor_normal(layer.ffn_norm, &rnd);
randomize_tensor_normal(layer.w1, &rnd);
randomize_tensor_normal(layer.w2, &rnd);
randomize_tensor_normal(layer.w3, &rnd);
}
}
void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
GGML_ASSERT(tensor->n_dims == 1);
GGML_ASSERT(tensor->ne[0] == ne0);
}
void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
GGML_ASSERT(tensor->n_dims == 2);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
}
void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
GGML_ASSERT(tensor->n_dims == 3);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
GGML_ASSERT(tensor->ne[2] == ne2);
}
void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
GGML_ASSERT(tensor->n_dims == 4);
GGML_ASSERT(tensor->ne[0] == ne0);
GGML_ASSERT(tensor->ne[1] == ne1);
GGML_ASSERT(tensor->ne[2] == ne2);
GGML_ASSERT(tensor->ne[3] == ne3);
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static size_t hash(void * p) {
return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static size_t hash_find(void * hash_table[], void * p) {
size_t h = hash(p);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// linear probing
size_t i = h;
while (hash_table[i] != NULL && hash_table[i] != p) {
i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
if (i == h) {
// visited all hash table entries -> not found
return GGML_GRAPH_HASHTABLE_SIZE;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return i;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static bool hash_insert(void * hash_table[], void * p) {
//size_t h = hash(p);
size_t i = hash_find(hash_table, p);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (hash_table[i] == p) {
return true;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// insert
GGML_ASSERT(hash_table[i] == NULL);
hash_table[i] = p;
return false;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static bool hash_contains(void * hash_table[], void * p) {
size_t i = hash_find(hash_table, p);
return (i < GGML_GRAPH_HASHTABLE_SIZE) && (hash_table[i] == p);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct hash_map {
void * keys[GGML_GRAPH_HASHTABLE_SIZE];
void * vals[GGML_GRAPH_HASHTABLE_SIZE];
};
//static const size_t HASH_MAP_SIZE = sizeof(struct hash_map);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct hash_map * new_hash_map() {
struct hash_map * result = new struct hash_map;
for (int i=0; i<GGML_GRAPH_HASHTABLE_SIZE; ++i) {
result->keys[i] = NULL;
result->vals[i] = NULL;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return result;
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void free_hash_map(struct hash_map * map) {
delete map;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static bool ggml_is_view(struct ggml_tensor * t) {
return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
switch (t->op) {
case GGML_OP_PERMUTE:
case GGML_OP_RESHAPE:
case GGML_OP_TRANSPOSE:
case GGML_OP_VIEW:
return t->src[0];
case GGML_OP_CPY:
return t->src[1];
default:
return NULL;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
struct ggml_tensor * parent = t;
do {
parent = get_view_parent(parent);
} while (ggml_is_view(parent));
return parent;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * ggml_recompute_graph_node(
struct ggml_context * ctx,
struct ggml_cgraph * graph,
struct hash_map * replacements,
struct ggml_tensor * node) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (node == NULL) {
return NULL;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (node->is_param) {
return node;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (!hash_contains(graph->visited_hash_table, node)) {
return node;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int count_children = 0;
for (int k = 0; k < GGML_MAX_SRC; ++k) {
if (node->src[k]) {
++count_children;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (count_children == 0) {
return node;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
size_t i = hash_find(replacements->keys, node);
GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
if (replacements->keys[i] == node) {
return (struct ggml_tensor *) replacements->vals[i];
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// insert clone into replacements
GGML_ASSERT(replacements->keys[i] == NULL); // assert that we don't overwrite
replacements->keys[i] = node;
replacements->vals[i] = clone;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
clone->op = node->op;
clone->grad = node->grad;
clone->is_param = node->is_param;
clone->extra = node->extra;
for (int k = 0; k < GGML_MAX_DIMS; ++k) {
clone->nb[k] = node->nb[k];
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
for (int k = 0; k < GGML_MAX_SRC; ++k) {
clone->src[k] = ggml_recompute_graph_node(ctx, graph, replacements, node->src[k]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (ggml_is_view(clone)) {
struct ggml_tensor * source = get_view_source(clone);
GGML_ASSERT(source != NULL);
clone->data = source->data;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(sizeof(node->op_params) == sizeof(int32_t) * (GGML_MAX_OP_PARAMS / sizeof(int32_t)));
GGML_ASSERT(sizeof(node->name) == GGML_MAX_NAME);
memcpy(clone->op_params, node->op_params, sizeof(node->op_params));
ggml_format_name(clone, "%s (clone)", ggml_get_name(node));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return clone;
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints) {
*gb_tmp = *gf;
ggml_build_backward_expand(ctx, gf, gb_tmp, true);
if (n_checkpoints <= 0) {
*gb = *gb_tmp;
return;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct hash_map * replacements = new_hash_map();
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// insert checkpoints in replacements
for (int i = 0; i < n_checkpoints; ++i) {
size_t k = hash_find(replacements->keys, checkpoints[i]);
GGML_ASSERT(k < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
GGML_ASSERT(replacements->keys[k] == NULL); // assert that we don't overwrite
replacements->keys[k] = checkpoints[i];
replacements->vals[k] = checkpoints[i];
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
*gb = *gf;
// rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes],
// replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]),
// by recomputing them from checkpoints
for (int i = gf->n_nodes; i<gb_tmp->n_nodes; ++i) {
struct ggml_tensor * node = gb_tmp->nodes[i];
for (int k = 0; k < GGML_MAX_SRC; ++k) {
// insert new tensors recomputing src, reusing already made replacements,
// remember replacements: remember new tensors with mapping from corresponding gf nodes
// recurse for input tensors,
// unless (i.e. terminating when) input tensors are checkpoints
node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// insert rewritten backward node with replacements made into resulting backward graph gb
ggml_build_forward_expand(gb, node);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
free_hash_map(replacements);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * llama_build_train_graphs(
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
struct my_llama_model * model,
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_allocr * alloc,
struct ggml_context * ctx,
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_cgraph * gb_tmp,
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
struct ggml_tensor * * logits,
struct ggml_tensor * tokens_input,
struct ggml_tensor * targets,
const int n_tokens,
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int n_batch,
const bool enable_flash_attn,
const bool enable_checkpointing) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_scratch(ctx, { 0, 0, nullptr, });
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
const int n_past = 0;
const int N = n_tokens;
const auto & hparams = model->hparams;
const int n_ctx = hparams.n_ctx;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
const int n_vocab = hparams.n_vocab;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_head = hparams.n_head;
const int n_rot = hparams.n_rot;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int n_ff = hparams.n_ff;
const float f_norm_rms_eps = hparams.f_norm_rms_eps;
const float rope_freq_base = hparams.rope_freq_base;
const float rope_freq_scale = hparams.rope_freq_scale;
auto set_name = [](struct ggml_tensor * t, const char * n) {
ggml_set_name(t, n);
if (t->grad) {
ggml_format_name(t->grad, "%s->grad", n);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
};
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// rope has so much parameters that we make a custom function for it
auto rope = [ctx, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
(struct ggml_tensor * t) -> struct ggml_tensor * {
// not capturing these, to silcence warnings
const int n_past = 0;
const int rope_mode = 0;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return ggml_rope_custom(ctx,
t, n_past, n_rot, rope_mode, n_ctx,
rope_freq_base, rope_freq_scale);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
};
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
set_name(tokens_input, "tokens_input");
set_name(targets, "targets");
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch); set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch);
struct ggml_tensor * t01 = ggml_get_rows(ctx, model->tok_embeddings, t00); set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * cur = t01;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<struct ggml_tensor *> checkpoints;
checkpoints.push_back(tokens_input);
checkpoints.push_back(targets);
checkpoints.push_back(t00);
checkpoints.push_back(t01);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * kv_scale;
if (!enable_flash_attn) {
kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_tensor * t02 = ggml_rms_norm (ctx, cur, f_norm_rms_eps); set_name(t02, "t02"); assert_shape_2d(t02, n_embd, N*n_batch);
struct ggml_tensor * t03 = ggml_repeat (ctx, layer.attention_norm, t02); set_name(t03, "t03"); assert_shape_2d(t03, n_embd, N*n_batch);
struct ggml_tensor * t04 = ggml_mul (ctx, t03, t02); set_name(t04, "t04"); assert_shape_2d(t04, n_embd, N*n_batch);
struct ggml_tensor * t05 = ggml_mul_mat (ctx, layer.wq, t04); set_name(t05, "t05"); assert_shape_2d(t05, n_embd, N*n_batch);
struct ggml_tensor * t06 = ggml_reshape_4d (ctx, t05, n_embd/n_head, n_head, N, n_batch); set_name(t06, "t06"); assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t07 = rope (t06); set_name(t07, "t07"); assert_shape_4d(t07, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t08 = ggml_mul_mat (ctx, layer.wk, t04); set_name(t08, "t08"); assert_shape_2d(t08, n_embd, N*n_batch);
struct ggml_tensor * t09 = ggml_reshape_4d (ctx, t08, n_embd/n_head, n_head, N, n_batch); set_name(t09, "t09"); assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t10 = rope (t09); set_name(t10, "t10"); assert_shape_4d(t10, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t11 = ggml_mul_mat (ctx, t04, layer.wv); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd);
struct ggml_tensor * t12 = ggml_reshape_4d (ctx, t11, N, n_batch, n_embd/n_head, n_head); set_name(t12, "t12"); assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head);
struct ggml_tensor * t13 = ggml_permute (ctx, t07, 0, 2, 1, 3); set_name(t13, "t13"); assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch);
struct ggml_tensor * t14 = ggml_permute (ctx, t10, 0, 2, 1, 3); set_name(t14, "t14"); assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch);
struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
struct ggml_tensor * t16;
if (enable_flash_attn) {
t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
} else {
struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past); set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch);
struct ggml_tensor * t16_3 = ggml_soft_max_inplace (ctx, t16_2); set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch);
t16 = ggml_mul_mat(ctx, t15, t16_3); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
}
struct ggml_tensor * t17 = ggml_permute (ctx, t16, 0, 2, 1, 3); set_name(t17, "t17"); assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t18 = ggml_cont (ctx, t17); set_name(t18, "t18"); assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch);
struct ggml_tensor * t19 = ggml_reshape_2d (ctx, t18, n_embd, N*n_batch); set_name(t19, "t19"); assert_shape_2d(t19, n_embd, N*n_batch);
struct ggml_tensor * t20 = ggml_mul_mat (ctx, layer.wo, t19); set_name(t20, "t20"); assert_shape_2d(t20, n_embd, N*n_batch);
struct ggml_tensor * t21 = ggml_add (ctx, t20, cur); set_name(t21, "t21"); assert_shape_2d(t21, n_embd, N*n_batch);
struct ggml_tensor * t22 = ggml_rms_norm (ctx, t21, f_norm_rms_eps); set_name(t22, "t22"); assert_shape_2d(t22, n_embd, N*n_batch);
struct ggml_tensor * t23 = ggml_repeat (ctx, layer.ffn_norm, t22); set_name(t23, "t23"); assert_shape_2d(t23, n_embd, N*n_batch);
struct ggml_tensor * t24 = ggml_mul (ctx, t23, t22); set_name(t24, "t24"); assert_shape_2d(t24, n_embd, N*n_batch);
struct ggml_tensor * t25 = ggml_mul_mat (ctx, layer.w3, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch);
struct ggml_tensor * t26 = ggml_mul_mat (ctx, layer.w1, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, N*n_batch);
struct ggml_tensor * t27 = ggml_silu (ctx, t26); set_name(t27, "t27"); assert_shape_2d(t27, n_ff, N*n_batch);
struct ggml_tensor * t28 = ggml_mul (ctx, t27, t25); set_name(t28, "t28"); assert_shape_2d(t28, n_ff, N*n_batch);
struct ggml_tensor * t29 = ggml_mul_mat (ctx, layer.w2, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, N*n_batch);
struct ggml_tensor * t30 = ggml_add (ctx, t29, t21); set_name(t30, "t30"); assert_shape_2d(t30, n_embd, N*n_batch);
cur = t30;
checkpoints.push_back(cur);
}
struct ggml_tensor * t31 = ggml_rms_norm (ctx, cur, f_norm_rms_eps); set_name(t31, "t31"); assert_shape_2d(t31, n_embd, N*n_batch);
struct ggml_tensor * t32 = ggml_repeat (ctx, model->norm, t31); set_name(t32, "t32"); assert_shape_2d(t32, n_embd, N*n_batch);
struct ggml_tensor * t33 = ggml_mul (ctx, t32, t31); set_name(t33, "t33"); assert_shape_2d(t33, n_embd, N*n_batch);
struct ggml_tensor * t34 = ggml_mul_mat (ctx, model->output, t33); set_name(t34, "t34"); assert_shape_2d(t34, n_vocab, N*n_batch);
struct ggml_tensor * t35 = ggml_reshape_3d (ctx, t34, n_vocab, N, n_batch); set_name(t35, "t35"); assert_shape_3d(t35, n_vocab, N, n_batch);
struct ggml_tensor * t36 = ggml_cross_entropy_loss(ctx, t35, targets); set_name(t36, "t36"); assert_shape_1d(t36, 1);
checkpoints.push_back(t31);
checkpoints.push_back(t32);
checkpoints.push_back(t33);
checkpoints.push_back(t34);
checkpoints.push_back(t35);
checkpoints.push_back(t36);
ggml_build_forward_expand(gf, t36);
if (enable_checkpointing) {
ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
} else {
*gb = *gf;
ggml_build_backward_expand(ctx, gf, gb, true);
}
if (alloc) {
// make sure some tensors are not reallocated by inserting new temporary nodes depending on them
int n_leafs_before = gb->n_leafs;
int n_nodes_before = gb->n_nodes;
struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
// output tensors
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
// input gradient
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
GGML_ASSERT(t36->grad->data == NULL && !ggml_is_view(t36->grad));
ggml_allocr_alloc(alloc, t36->grad);
// gradient tensors (will be set to zero by ggml_graph_reset)
// pinning these produces large unnecessary memory overhead, which will be resolved by PR 2632
for (int i = 0; i < gf->n_nodes; ++i) {
if (!gf->grads[i]) continue;
if (gf->grads[i]->data == NULL && !ggml_is_view(gf->grads[i])) {
ggml_allocr_alloc(alloc, gf->grads[i]);
}
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, gf->grads[i], one));
}
// allocating checkpoints in one block to reduce memory fragmentation
// note: they will be freed in reverse order
for (int i = 0; i < (int) checkpoints.size(); ++i) {
if (checkpoints[i]->data == NULL && !ggml_is_view(checkpoints[i])) {
ggml_allocr_alloc(alloc, checkpoints[i]);
}
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
//int n_leafs_after = gb->n_leafs;
//int n_nodes_after = gb->n_nodes;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_allocr_alloc_graph(alloc, gb);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// remove the additional nodes and leafs
for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
gb->leafs[i] = NULL;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
for (int i = n_nodes_before; i < gb->n_nodes; ++i) {
gb->nodes[i] = NULL;
}
gb->n_leafs = n_leafs_before;
gb->n_nodes = n_nodes_before;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
*logits = t35;
return t36;
}
void set_f32_3d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, int64_t i2, float value) {
float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
*ptr = value;
}
void set_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, float value) {
float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
*ptr = value;
}
void set_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, int32_t value) {
int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
*ptr = value;
}
float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
return *ptr;
}
int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
return *ptr;
}
void print_row(struct ggml_tensor * probs, int i) {
for (int k = 0; k < probs->ne[0]; ++k) {
float p = get_f32_2d(probs, k, i);
printf(" %.2f", p);
}
printf("\n");
}
void print_matrix(struct ggml_tensor * probs) {
assert(probs->n_dims == 2);
for (int i = 0; i < probs->ne[1]; ++i) {
for (int k = 0; k < probs->ne[0]; ++k) {
float p = get_f32_2d(probs, k, i);
printf(" %.2f", p);
}
printf("\n");
}
}
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
void get_example_targets(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int n_tokens = tokens_input->ne[0];
int n_vocab = target_logits->ne[0];
size_t sample = train_samples[example_id % n_train_samples];
GGML_ASSERT(sample+n_tokens-1 < n_train_data);
ggml_set_f32(target_logits, -1.0f/n_vocab);
ggml_set_f32(target_probs, 0.0f);
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
ggml_set_i32_1d(tokens_input, 0, llama_token_bos(lctx));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
for (int i=1; i<n_tokens+1; ++i) {
int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
set_f32_2d(target_logits, token, i-1, +1.0f);
set_f32_2d(target_probs, token, i-1, +1.0f);
if (i<n_tokens) {
ggml_set_i32_1d(tokens_input, i, token);
}
}
}
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
void get_example_targets_batch(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
GGML_ASSERT(tokens_input->n_dims == 2);
GGML_ASSERT(target_logits->n_dims == 3);
GGML_ASSERT(target_probs->n_dims == 3);
int n_vocab = target_logits->ne[0];
int n_tokens = tokens_input->ne[0];
int n_batch = tokens_input->ne[1];
GGML_ASSERT(n_tokens == target_logits->ne[1]);
GGML_ASSERT(n_batch == target_logits->ne[2]);
GGML_ASSERT(n_vocab == target_probs->ne[0]);
GGML_ASSERT(n_tokens == target_probs->ne[1]);
GGML_ASSERT(n_batch == target_probs->ne[2]);
ggml_set_f32(target_logits, -1.0f/n_vocab);
ggml_set_f32(target_probs, 0.0f);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
for (int k=0; k<n_batch; ++k) {
// printf("%s: batch %d\n", __func__, k);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
size_t sample_idx = (example_id*n_batch + k) % n_train_samples;
size_t sample = train_samples[sample_idx];
// printf("%s: sample_idx=%zu sample=%zu\n", __func__, sample_idx, sample);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
GGML_ASSERT(sample+n_tokens-1 < n_train_data);
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
set_i32_2d(tokens_input, 0, k, llama_token_bos(lctx));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
for (int i=1; i<n_tokens+1; ++i) {
int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
set_f32_3d(target_logits, token, i-1, k, +1.0f);
set_f32_3d(target_probs, token, i-1, k, +1.0f);
if (i<n_tokens) {
set_i32_2d(tokens_input, i, k, token);
}
}
}
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int tokenize_file(struct llama_context * lctx, const char * filename, std::vector<llama_token>& out) {
FILE * fp = std::fopen(filename, "rb");
if (fp == NULL) {
return 0;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
#ifdef _WIN32
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(_fseeki64(fp, (__int64) 0, SEEK_END) == 0);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#else
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(std::fseek(fp, (long) 0, SEEK_END) == 0);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#endif
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
size_t size = 0;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#ifdef _WIN32
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
__int64 ret = _ftelli64(fp);
size = ret;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#else
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
long ret = std::ftell(fp);
size = ret;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
#endif
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
#ifdef _WIN32
GGML_ASSERT(_fseeki64(fp, (__int64) 0, SEEK_SET) == 0);
#else
GGML_ASSERT(std::fseek(fp, (long) 0, SEEK_SET) == 0);
#endif
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<char> buf;
buf.resize(size+1);
out.resize(size+1);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (std::fread(buf.data(), size, 1, fp) != 1) {
die("unexpectedly reached end of file");
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (ferror(fp)) {
die_fmt("fread failed: %s", strerror(errno));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
buf[size] = '\0';
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int n_tokens = llama_tokenize(lctx, buf.data(), buf.size(), out.data(), out.size(), false);
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
if (n_tokens < 0) {
out.resize(-n_tokens);
n_tokens = llama_tokenize(lctx, buf.data(), buf.size(), out.data(), out.size(), false);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGML_ASSERT(n_tokens >= 0);
out.resize(n_tokens);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
bool verify = false;
if (verify) {
const char * in = buf.data();
const char * end = buf.data() + buf.size();
for (int i = 0; i < (int) out.size(); ++i) {
std::string s = llama_token_to_piece(lctx, out[i]);
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
int len = s.length();
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (in >= end) {
printf("%s: unexpected end of original text.\n", __func__);
break;
}
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
const bool matches = (strncmp(in, s.c_str(), len) == 0);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (matches) {
in += len;
} else {
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
printf("%s: mismatch: expected '%s', but got '%s'\n", __func__, std::string(in, len).c_str(), s.c_str());
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
}
return n_tokens;
}
void shuffle_ints(int * begin, int * end) {
if (end <= begin) return;
int max=begin[0];
for (int i=1; i<end-begin; ++i) {
if (begin[i] > max) {
max = begin[i];
}
}
std::vector<float> vals;
vals.resize(max+1);
for (int i=0; i<max+1; ++i) {
vals[i] = frand();
}
std::sort(begin, end, [&vals](int a, int b){
return vals.at(a) < vals.at(b);
});
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
{ \
const std::string skey(key); \
const int kid = gguf_find_key(ctx, skey.c_str()); \
if (kid >= 0) { \
enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
if (ktype != (type)) { \
die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} \
(dst) = func(ctx, kid); \
} else if (req) { \
die_fmt("key not found in model: %s", skey.c_str()); \
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} \
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
bool are_same_layout(struct ggml_tensor * a, struct ggml_tensor * b) {
GGML_ASSERT(a != NULL);
GGML_ASSERT(b != NULL);
GGML_ASSERT(a->type == b->type);
GGML_ASSERT(ggml_are_same_shape(a, b));
GGML_ASSERT(ggml_is_contiguous(a) && ggml_is_contiguous(b));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return true;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void read_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name) {
if (dst == NULL) {
return;
}
struct ggml_tensor * t = ggml_get_tensor(ctx, name);
GGML_ASSERT(are_same_layout(dst, t));
memcpy(dst->data, t->data, ggml_nbytes(t));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (strlen(ggml_get_name(dst)) == 0) {
ggml_set_name(dst, name);
}
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt) {
// NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
uint32_t file_version;
GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_FILE_VERSION);
GGML_ASSERT(file_version == 0);
GGUF_GET_KEY(fctx, opt->params.past, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT);
GGUF_GET_KEY(fctx, opt->iter, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ITERATION_COUNT);
GGUF_GET_KEY(fctx, opt->just_initialized, gguf_get_val_bool, GGUF_TYPE_BOOL, true, LLM_KV_OPTIMIZER_JUST_INITIALIZED);
uint64_t nx;
GGUF_GET_KEY(fctx, nx, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_OPTIMIZER_PARAMETER_COUNT);
opt->nx = (size_t) nx;
// don't call ggml_opt_init until optimizer type and optimizer specific parameters are know
std::string opt_type;
GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
opt->params.type = GGML_OPT_ADAM;
GGUF_GET_KEY(fctx, opt->adam.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
GGUF_GET_KEY(fctx, opt->adam.fx_prev, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
GGUF_GET_KEY(fctx, opt->adam.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT);
GGML_ASSERT(opt->ctx != NULL);
ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
read_tensor_by_name(opt->adam.m, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
read_tensor_by_name(opt->adam.v, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
read_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
} else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
opt->params.type = GGML_OPT_LBFGS;
GGUF_GET_KEY(fctx, opt->params.lbfgs.m, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
GGUF_GET_KEY(fctx, opt->lbfgs.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
GGUF_GET_KEY(fctx, opt->lbfgs.step, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP);
GGUF_GET_KEY(fctx, opt->lbfgs.j, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J);
GGUF_GET_KEY(fctx, opt->lbfgs.k, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K);
GGUF_GET_KEY(fctx, opt->lbfgs.end, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END);
GGUF_GET_KEY(fctx, opt->lbfgs.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT);
GGML_ASSERT(opt->ctx != NULL);
ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
read_tensor_by_name(opt->lbfgs.x, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
read_tensor_by_name(opt->lbfgs.xp, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
read_tensor_by_name(opt->lbfgs.g, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
read_tensor_by_name(opt->lbfgs.gp, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
read_tensor_by_name(opt->lbfgs.d, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
read_tensor_by_name(opt->lbfgs.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
read_tensor_by_name(opt->lbfgs.lmal, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
read_tensor_by_name(opt->lbfgs.lmys, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
read_tensor_by_name(opt->lbfgs.lms, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
read_tensor_by_name(opt->lbfgs.lmy, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
} else {
die("unknown optimizer type");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt) {
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_FILE_VERSION, 0);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, opt->params.past);
gguf_set_val_u64(fctx, LLM_KV_OPTIMIZER_PARAMETER_COUNT, (uint64_t) opt->nx);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ITERATION_COUNT, opt->iter);
gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
switch (opt->params.type) {
case GGML_OPT_ADAM:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, opt->adam.fx_best);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, opt->adam.fx_prev);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, opt->adam.n_no_improvement);
ggml_set_name(opt->adam.m, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
ggml_set_name(opt->adam.v, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
if (opt->adam.pf) {
ggml_set_name(opt->adam.pf, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
gguf_add_tensor(fctx, opt->adam.m);
gguf_add_tensor(fctx, opt->adam.v);
if (opt->adam.pf) {
gguf_add_tensor(fctx, opt->adam.pf);
}
} break;
case GGML_OPT_LBFGS:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, opt->lbfgs.fx_best);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, opt->lbfgs.step);
gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, opt->lbfgs.j);
gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, opt->lbfgs.k);
gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, opt->lbfgs.end);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, opt->lbfgs.n_no_improvement);
ggml_set_name(opt->lbfgs.x, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
ggml_set_name(opt->lbfgs.xp, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
ggml_set_name(opt->lbfgs.g, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
ggml_set_name(opt->lbfgs.gp, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
ggml_set_name(opt->lbfgs.d, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
if (opt->lbfgs.pf) {
ggml_set_name(opt->lbfgs.pf, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
}
ggml_set_name(opt->lbfgs.lmal, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
ggml_set_name(opt->lbfgs.lmys, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
ggml_set_name(opt->lbfgs.lms, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
ggml_set_name(opt->lbfgs.lmy, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
gguf_add_tensor(fctx, opt->lbfgs.x);
gguf_add_tensor(fctx, opt->lbfgs.xp);
gguf_add_tensor(fctx, opt->lbfgs.g);
gguf_add_tensor(fctx, opt->lbfgs.gp);
gguf_add_tensor(fctx, opt->lbfgs.d);
if (opt->lbfgs.pf) {
gguf_add_tensor(fctx, opt->lbfgs.pf);
}
gguf_add_tensor(fctx, opt->lbfgs.lmal);
gguf_add_tensor(fctx, opt->lbfgs.lmys);
gguf_add_tensor(fctx, opt->lbfgs.lms);
gguf_add_tensor(fctx, opt->lbfgs.lmy);
} break;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model) {
// NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
std::string arch;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<char> keybuf;
keybuf.resize(512);
auto kv = [&arch, &keybuf](const char * key) -> const char * {
snprintf(keybuf.data(), keybuf.size(), key, arch.c_str());
return keybuf.data();
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<char> tn_buf;
tn_buf.resize(GGML_MAX_NAME);
auto tn = [&tn_buf](const char * key) -> const char * {
snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
return tn_buf.data();
};
auto tni = [&tn_buf](const char * key, int bid) -> const char * {
snprintf(tn_buf.data(), tn_buf.size(), key, bid);
std::string s = tn_buf.data();
snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
return tn_buf.data();
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
GGML_ASSERT(arch == "llama");
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
uint32_t ftype_u;
GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE);
GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// n_ctx was not saved in earlier checkpoint file versions, so we make it optional here
GGUF_GET_KEY(fctx, model->hparams.n_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGUF_GET_KEY(fctx, model->hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH));
GGUF_GET_KEY(fctx, model->hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH));
GGUF_GET_KEY(fctx, model->hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
GGUF_GET_KEY(fctx, model->hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
model->hparams.n_rot = model->hparams.n_embd / model->hparams.n_head;
GGUF_GET_KEY(fctx, model->hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float rope_freq_scale = 1.0f;
GGUF_GET_KEY(fctx, model->hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
GGUF_GET_KEY(fctx, model->hparams.rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
GGUF_GET_KEY(fctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
if (rope_freq_scale != 1.0f) {
model->hparams.rope_freq_scale = 1.0f / rope_freq_scale;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
init_model(model);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
read_tensor_by_name(model->tok_embeddings, f_ggml_ctx, tn(LLM_TENSOR_TOKEN_EMBD));
read_tensor_by_name(model->norm, f_ggml_ctx, tn(LLM_TENSOR_OUTPUT_NORM));
read_tensor_by_name(model->output, f_ggml_ctx, tn(LLM_TENSOR_OUTPUT));
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
auto & layer = model->layers[i];
read_tensor_by_name(layer.attention_norm, f_ggml_ctx, tni(LLM_TENSOR_ATTN_NORM, i));
read_tensor_by_name(layer.wq, f_ggml_ctx, tni(LLM_TENSOR_ATTN_Q, i));
read_tensor_by_name(layer.wk, f_ggml_ctx, tni(LLM_TENSOR_ATTN_K, i));
read_tensor_by_name(layer.wv, f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
read_tensor_by_name(layer.wo, f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
read_tensor_by_name(layer.ffn_norm, f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
read_tensor_by_name(layer.w1, f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
read_tensor_by_name(layer.w2, f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
read_tensor_by_name(layer.w3, f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model) {
const char * arch = "llama";
enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
std::vector<char> keybuf;
keybuf.resize(512);
auto kv = [arch, &keybuf](const char * key) -> const char * {
snprintf(keybuf.data(), keybuf.size(), key, arch);
return keybuf.data();
};
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// set arch
gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// set hparams
gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH), model->hparams.n_ctx );
gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH), model->hparams.n_embd );
gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH), model->hparams.n_ff );
gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT), model->hparams.n_head );
gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT), model->hparams.n_layer );
gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT), model->hparams.n_rot );
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps );
gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE), model->hparams.rope_freq_base ); // TODO load in llama.cpp
gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR), 1.0f / model->hparams.rope_freq_scale );
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// set vocab by copying from vocab_model gguf file
{
struct gguf_init_params params = {
/*.no_alloc = */ false,
/*.ctx = */ NULL,
};
struct gguf_context * vctx = gguf_init_from_file(fn_vocab_model, params);
const int token_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_LIST));
if (token_idx == -1) {
die("cannot find tokenizer vocab in model file");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
const uint32_t n_vocab = gguf_get_arr_n(vctx, token_idx);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int score_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_SCORES));
if (score_idx == -1) {
die("cannot find tokenizer scores in model file");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const float * scores = (const float * ) gguf_get_arr_data(vctx, score_idx);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int toktype_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE));
if (toktype_idx == -1) {
die("cannot find token type list in GGUF file");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int * toktypes = (const int * ) gguf_get_arr_data(vctx, toktype_idx);
std::string tokenizer_name;
GGUF_GET_KEY(vctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL));
gguf_set_val_str(fctx, kv(LLM_KV_TOKENIZER_MODEL), tokenizer_name.c_str());
gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_SCORES), GGUF_TYPE_FLOAT32, scores, n_vocab);
gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE), GGUF_TYPE_INT32, toktypes, n_vocab);
int32_t special_bos_id = 1;
int32_t special_eos_id = 2;
int32_t special_unk_id = 0;
int32_t special_sep_id = -1;
int32_t special_pad_id = -1;
if (tokenizer_name == "llama") {
// default special tokens
special_bos_id = 1;
special_eos_id = 2;
special_unk_id = 0;
special_sep_id = -1;
special_pad_id = -1;
} else if (tokenizer_name == "gpt2") {
// read and copy bpe merges
const int merges_keyidx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_MERGES));
if (merges_keyidx == -1) {
die("cannot find tokenizer merges in model file");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const int n_merges = gguf_get_arr_n(vctx, merges_keyidx);
std::vector<const char*> merges;
merges.resize(n_merges);
for (int i = 0; i < n_merges; i++) {
merges[i] = gguf_get_arr_str(vctx, merges_keyidx, i);
}
gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_MERGES), merges.data(), n_merges);
// default special tokens
special_bos_id = 11;
special_eos_id = 11;
special_unk_id = -1;
special_sep_id = -1;
special_pad_id = -1;
} else {
fprintf(stderr, "%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str());
fprintf(stderr, "%s: using default tokenizer: 'llama'", __func__);
}
std::vector<const char*> tokens;
tokens.resize(n_vocab);
for (uint32_t i = 0; i < n_vocab; i++) {
tokens[i] = gguf_get_arr_str(vctx, token_idx, i);
}
gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_LIST), tokens.data(), n_vocab);
GGUF_GET_KEY(vctx, special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
GGUF_GET_KEY(vctx, special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
GGUF_GET_KEY(vctx, special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
GGUF_GET_KEY(vctx, special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
GGUF_GET_KEY(vctx, special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_BOS_ID), special_bos_id);
gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_EOS_ID), special_eos_id);
gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_UNK_ID), special_unk_id);
gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_SEP_ID), special_sep_id);
gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_PAD_ID), special_pad_id);
gguf_free(vctx);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// add tensors
gguf_add_tensor(fctx, model->tok_embeddings);
gguf_add_tensor(fctx, model->norm);
gguf_add_tensor(fctx, model->output);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
auto & layer = model->layers[i];
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
gguf_add_tensor(fctx, layer.attention_norm);
gguf_add_tensor(fctx, layer.wq);
gguf_add_tensor(fctx, layer.wk);
gguf_add_tensor(fctx, layer.wv);
gguf_add_tensor(fctx, layer.wo);
gguf_add_tensor(fctx, layer.ffn_norm);
gguf_add_tensor(fctx, layer.w1);
gguf_add_tensor(fctx, layer.w2);
gguf_add_tensor(fctx, layer.w3);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
}
void save_llama_model_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model) {
struct gguf_context * fctx = gguf_init_empty();
save_llama_model_gguf(fctx, fn_vocab_model, model);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// write file
const bool only_meta = false;
gguf_write_to_file(fctx, filename, only_meta);
gguf_free(fctx);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct ggml_opt_context * opt) {
load_llama_model_gguf(fctx, f_ggml_ctx, model);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
uint32_t file_version;
GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
GGML_ASSERT(file_version == 0);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
GGUF_GET_KEY(fctx, model->train_its, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT);
GGUF_GET_KEY(fctx, model->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT);
GGUF_GET_KEY(fctx, model->train_tokens, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
load_opt_context_gguf(fctx, f_ggml_ctx, opt);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void save_checkpoint_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) {
save_llama_model_gguf(fctx, fn_vocab_model, model);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
gguf_set_val_u32(fctx, LLM_KV_TRAINING_FILE_VERSION, 0);
gguf_set_val_u32(fctx, LLM_KV_TRAINING_ITERATION_COUNT, model->train_its);
gguf_set_val_u32(fctx, LLM_KV_TRAINING_SAMPLE_COUNT, model->train_samples);
gguf_set_val_u32(fctx, LLM_KV_TRAINING_TOKEN_COUNT, model->train_tokens);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
save_opt_context_gguf(fctx, opt);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
bool load_checkpoint_file(const char * filename, struct my_llama_model * model, struct ggml_opt_context * opt) {
struct ggml_context * f_ggml_ctx;
struct gguf_init_params params;
params.no_alloc = false;
params.ctx = &f_ggml_ctx;
struct gguf_context * fctx = gguf_init_from_file(filename, params);
if (fctx == NULL) {
return false;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
load_checkpoint_gguf(fctx, f_ggml_ctx, model, opt);
return true;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
void save_checkpoint_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) {
struct gguf_context * fctx = gguf_init_empty();
save_checkpoint_gguf(fctx, fn_vocab_model, model, opt);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
// write file
const bool only_meta = false;
gguf_write_to_file(fctx, filename, only_meta);
gguf_free(fctx);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float cosine_decay(const int decay_steps, const float minimum, int step) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (step > decay_steps) {
step = decay_steps;
}
const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps));
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
const float decay = (1 - minimum)*cosine_decay + minimum;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
return decay;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float cosine_decay_restart(int decay_steps, const float minimum, int step, float restart_step_mult, bool enable_restart) {
if (enable_restart) {
while (step > decay_steps) {
step -= decay_steps;
decay_steps = (int) restart_step_mult * decay_steps;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
return cosine_decay(decay_steps, minimum, step);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
struct train_params {
const char * fn_vocab_model;
const char * fn_train_data;
const char * fn_checkpoint_in;
const char * fn_checkpoint_out;
const char * fn_model_out;
2023-07-01 17:45:44 +02:00
uint32_t seed;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int n_ctx;
int n_embd;
int n_head;
int n_layer;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int n_ff;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int n_threads;
int n_batch;
int n_examples;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float f_norm_rms_eps;
float rope_freq_base;
float rope_freq_scale;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int print_info_interval;
bool samples_start_after_nl;
bool use_adam;
bool use_flash;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
bool use_checkpointing;
bool use_alloc;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
// only adam
int warmup;
int cos_decay_steps;
float cos_decay_restart;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float cos_decay_min;
bool enable_restart;
int opt_past;
float opt_delta;
int opt_max_no_improvement;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int lbfgs_n_iter;
int adam_n_iter;
float adam_alpha;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
float adam_min_alpha;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
float adam_decay;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int adam_decay_min_ndim;
float adam_beta1;
float adam_beta2;
float adam_gclip;
float adam_eps_f;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int mem_model_gb;
int mem_compute_gb;
int mem_compute0_gb;
};
struct train_params get_default_train_params() {
struct train_params params;
params.fn_vocab_model = "ggml-vic7b-uncensored-q4_0.bin";
params.fn_train_data = "shakespeare.txt";
params.fn_checkpoint_in = "checkpoint.bin";
params.fn_checkpoint_out = "checkpoint.bin";
params.fn_model_out = "ggml-checkpoint-f32.bin";
params.seed = -1;
params.n_ctx = 128;
params.n_embd = 256;
params.n_head = 8;
params.n_layer = 16;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.n_ff = 768;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
params.n_threads = 6;
params.n_batch = 8;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.n_examples = 1;
params.f_norm_rms_eps = 1e-5;
params.rope_freq_base = 10000.0f;
params.rope_freq_scale = 1.0f;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
params.print_info_interval = 1;
params.samples_start_after_nl = false;
params.use_adam = true;
params.use_flash = true;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.use_checkpointing = true;
params.use_alloc = true;
params.opt_past = 0;
params.opt_delta = 1e-5f;
params.opt_max_no_improvement = 0;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
// only adam
params.warmup = 100;
params.cos_decay_steps = 1000;
params.cos_decay_restart = 1.1f;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.cos_decay_min = 0.1f;
params.enable_restart = false;
params.lbfgs_n_iter = 256;
params.adam_n_iter = 256;
params.adam_alpha = 1e-3f;
params.adam_min_alpha = 0;
params.adam_decay = 1e-1f;
params.adam_decay_min_ndim = 2;
params.adam_beta1 = 0.9f;
params.adam_beta2 = 0.999f;
params.adam_gclip = 1.0f;
params.adam_eps_f = 0.0f;
params.mem_model_gb = 2;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
params.mem_compute_gb = 24;
params.mem_compute0_gb = 8;
return params;
}
void train_print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " --vocab-model FNAME model path from which to load vocab (default '%s')\n", params->fn_vocab_model);
fprintf(stderr, " --train-data FNAME path from which to load training data (default '%s')\n", params->fn_train_data);
fprintf(stderr, " --checkpoint-in FNAME path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
fprintf(stderr, " --checkpoint-out FNAME path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
fprintf(stderr, " --model-out FNAME path to save ggml model (default '%s')\n", params->fn_model_out);
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for -1)\n");
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " -c N, --ctx N Context size used during training (default %d)\n", params->n_ctx);
fprintf(stderr, " --embd N Embedding size used for new models (default %d)\n", params->n_embd);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --ff N Feedforward size used for new models. (default %d)\n", params->n_ff);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " --head N Number of heads for new models (default %d)\n", params->n_head);
fprintf(stderr, " --layer N Number of layers for new models (default %d)\n", params->n_layer);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --norm-rms-eps F RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps);
fprintf(stderr, " --rope-freq-base F Frequency base for ROPE (default %f)\n", params->rope_freq_base);
fprintf(stderr, " --rope-freq-scale F Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " -t N, --threads N Number of threads (default %d)\n", params->n_threads);
fprintf(stderr, " -b N, --batch N Parallel batch size (default %d)\n", params->n_batch);
fprintf(stderr, " -n N, --examples N Number of examples to train (default %d)\n", params->n_examples);
fprintf(stderr, " --print-info-interval N Print infos during training each N examples (default %d)\n", params->print_info_interval);
fprintf(stderr, " --samples-after-nl Training samples start after newlines. (default %s)\n", params->samples_start_after_nl ? "on" : "off");
fprintf(stderr, " --use-lbfgs Use LBFGS optimizer instead of default Adam\n");
fprintf(stderr, " --use-adam Use Adam optimizer (default)\n");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --no-flash Don't use flash attention \n");
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " --use-flash Use flash attention (default)\n");
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --no-checkpointing Don't use gradient checkpointing\n");
fprintf(stderr, " --use-checkpointing Use gradient checkpointing (default)\n");
fprintf(stderr, " --no-alloc Don't use allocator\n");
fprintf(stderr, " --use-alloc Use allocator (default)\n");
fprintf(stderr, " --warmup N Only for Adam optimizer. Number of warmup steps (default %d)\n", params->warmup);
fprintf(stderr, " --cos-decay-steps N Only for Adam optimizer. Number of cosine decay steps (default %d)\n", params->cos_decay_steps);
fprintf(stderr, " --cos-decay-restart N Only for Adam optimizer. Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart);
fprintf(stderr, " --cos-decay-min N Only for Adam optimizer. Cosine decay minimum (default %f)\n", params->cos_decay_min);
fprintf(stderr, " --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay %s\n", params->enable_restart ? "(default)" : "");
fprintf(stderr, " --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay %s\n", !params->enable_restart ? "(default)" : "");
fprintf(stderr, " --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. (default %d)\n", params->opt_past);
fprintf(stderr, " --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. (default %f)\n", params->opt_delta);
fprintf(stderr, " --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. (default %d)\n", params->opt_max_no_improvement);
fprintf(stderr, " --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " --adam-iter N Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter);
fprintf(stderr, " --adam-alpha N Adam learning rate alpha (default %f)\n", params->adam_alpha);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --adam-min-alpha N Adam minimum learning rate alpha - including warmup phase (default %f)\n", params->adam_min_alpha);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " --adam-decay N AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --adam-decay-min-ndim N Minimum number of tensor dimensions to apply AdamW weight decay. Weight decay is not applied to tensors with less n_dims. (default %d)\n", params->adam_decay_min_ndim);
fprintf(stderr, " --adam-beta1 N AdamW beta1 in interval [0,1). How much to smooth the first moment of gradients. (default %f)\n", params->adam_beta1);
fprintf(stderr, " --adam-beta2 N AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
fprintf(stderr, " --adam-gclip N AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, " --mem-model N Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb);
fprintf(stderr, " --mem-compute N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
fprintf(stderr, " --mem-compute0 N Memory to allocate for automatic memory allocator in gigabytes. (default %d)\n", params->mem_compute0_gb);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
fprintf(stderr, "\n");
}
bool train_params_parse(int argc, char ** argv, struct train_params * params) {
bool invalid_param = false;
std::string arg;
struct train_params default_params = get_default_train_params();
const std::string arg_prefix = "--";
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
std::replace(arg.begin(), arg.end(), '_', '-');
}
if (arg == "--vocab-model") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_vocab_model = argv[i];
} else if (arg == "--train-data") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_train_data = argv[i];
} else if (arg == "--checkpoint-in") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_checkpoint_in = argv[i];
} else if (arg == "--checkpoint-out") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_checkpoint_out = argv[i];
} else if (arg == "--model-out") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_model_out = argv[i];
} else if (arg == "-s" || arg == "--seed") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->seed = std::stoi(argv[i]);
} else if (arg == "-c" || arg == "--ctx") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_ctx = std::stoi(argv[i]);
} else if (arg == "--embd") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_embd = std::stoi(argv[i]);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--ff") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->n_ff = std::stoi(argv[i]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--head") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_head = std::stoi(argv[i]);
} else if (arg == "--layer") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_layer = std::stoi(argv[i]);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--norm-rms-eps") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->f_norm_rms_eps = std::stof(argv[i]);
} else if (arg == "--rope-freq-base") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->rope_freq_base = std::stof(argv[i]);
} else if (arg == "--rope-freq-scale") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->rope_freq_scale = std::stof(argv[i]);
} else if (arg == "-t" || arg == "--threads") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->n_threads = std::stoi(argv[i]);
} else if (arg == "-b" || arg == "--batch") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->n_batch = std::stoi(argv[i]);
} else if (arg == "-n" || arg == "--examples") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->n_examples = std::stoi(argv[i]);
} else if (arg == "--print-info-interval") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->print_info_interval = std::stoi(argv[i]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--samples-after-nl") {
params->samples_start_after_nl = true;
} else if (arg == "--use-lbfgs") {
params->use_adam = false;
} else if (arg == "--use-adam") {
params->use_adam = true;
} else if (arg == "--no-flash") {
params->use_flash = false;
} else if (arg == "--use-flash") {
params->use_flash = true;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--no-checkpointing") {
params->use_checkpointing = false;
} else if (arg == "--use-checkpointing") {
params->use_checkpointing = true;
} else if (arg == "--no-alloc") {
params->use_alloc = false;
} else if (arg == "--use-alloc") {
params->use_alloc = true;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--warmup") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->warmup = std::stoi(argv[i]);
} else if (arg == "--cos-decay-steps") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->cos_decay_steps = std::stof(argv[i]);
} else if (arg == "--cos-decay-restart") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->cos_decay_restart = std::stof(argv[i]);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--cos-decay-min") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->cos_decay_min = std::stof(argv[i]);
} else if (arg == "--enable-restart") {
params->enable_restart = true;
} else if (arg == "--disable-restart") {
params->enable_restart = false;
} else if (arg == "--opt-past") {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (++i >= argc) {
invalid_param = true;
break;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params->opt_past = std::stoi(argv[i]);
} else if (arg == "--opt-delta") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->opt_delta = std::stof(argv[i]);
} else if (arg == "--opt-max-no-improvement") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->opt_max_no_improvement = std::stoi(argv[i]);
} else if (arg == "--adam-epsf") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_eps_f = std::stof(argv[i]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--adam-iter") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_n_iter = std::stoi(argv[i]);
} else if (arg == "--adam-alpha") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_alpha = std::stof(argv[i]);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--adam-min-alpha") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_min_alpha = std::stof(argv[i]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--adam-decay") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_decay = std::stof(argv[i]);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
} else if (arg == "--adam-decay-min-ndim") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_decay_min_ndim = std::stoi(argv[i]);
} else if (arg == "--adam-beta1") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_beta1 = std::stof(argv[i]);
} else if (arg == "--adam-beta2") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_beta2 = std::stof(argv[i]);
} else if (arg == "--adam-gclip") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->adam_gclip = std::stof(argv[i]);
} else if (arg == "--lbfgs-iter") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->lbfgs_n_iter = std::stoi(argv[i]);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
} else if (arg == "--mem-model") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->mem_model_gb = std::stoi(argv[i]);
} else if (arg == "--mem-compute") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->mem_compute_gb = std::stoi(argv[i]);
} else if (arg == "--mem-compute0") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->mem_compute0_gb = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") {
train_print_usage(argc, argv, &default_params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
train_print_usage(argc, argv, &default_params);
exit(1);
}
}
if (invalid_param) {
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
train_print_usage(argc, argv, &default_params);
exit(1);
}
return true;
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct opt_callback_data {
struct train_params * params;
struct ggml_opt_context * opt;
struct llama_context * lctx;
llama_token * tokens_data;
size_t tokens_size;
int * samples_data;
size_t samples_size;
int shuffle_countdown;
struct ggml_tensor * tokens_input;
struct ggml_tensor * target_logits;
struct ggml_tensor * target_probs;
};
void opt_callback(void * vdata, float * sched) {
struct opt_callback_data * data = (struct opt_callback_data *) vdata;
struct train_params * params = data->params;
struct ggml_opt_context * opt = data->opt;
int n_batch = params->n_batch;
*sched = (opt->iter < params->warmup)
? (float) opt->iter / (float) params->warmup
: cosine_decay_restart(
params->cos_decay_steps,
params->cos_decay_min,
opt->iter - params->warmup,
params->cos_decay_restart,
params->enable_restart);
float min_sched = params->adam_min_alpha / params->adam_alpha;
*sched = min_sched + *sched * (1.0f - min_sched);
int impr_plot = std::isnan(opt->loss_after) ? 0 : -std::lround(1 + (opt->loss_before - opt->loss_after) * 10.0f);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
printf("%s: iter=%*d, sched=%f loss0=%f loss=%f | improvement: %*d>\n", __func__, 6, opt->iter, *sched, opt->loss_before, opt->loss_after, impr_plot, (int)0);
if (data->shuffle_countdown < n_batch) {
printf("%s: reshuffle samples\n", __func__);
shuffle_ints(data->samples_data, data->samples_data + data->samples_size);
for (int i = 0; i < (int) data->samples_size; ++i) {
GGML_ASSERT(data->samples_data[i]+params->n_ctx-1 < (int) data->tokens_size);
}
data->shuffle_countdown = data->samples_size;
}
get_example_targets_batch(
data->lctx,
data->samples_data,
data->samples_size,
data->tokens_data,
data->tokens_size,
opt->iter,
data->tokens_input,
data->target_logits,
data->target_probs);
data->shuffle_countdown -= n_batch;
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
int main(int argc, char ** argv) {
struct train_params params = get_default_train_params();
if (!train_params_parse(argc, argv, &params)) {
return 1;
}
if (params.seed == LLAMA_DEFAULT_SEED) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
params.seed = time(NULL);
}
printf("%s: seed: %u\n", __func__, params.seed);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
srand(params.seed);
struct llama_context_params llama_params = llama_context_default_params();
llama_params.vocab_only = true;
struct llama_model * lmodel = llama_load_model_from_file(params.fn_vocab_model, llama_params);
struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
printf("%s: tokenize training data\n", __func__);
std::vector<llama_token> train_tokens;
if (tokenize_file(lctx, params.fn_train_data, train_tokens) < 0) {
fprintf(stderr, "%s: failed to tokenize file '%s'\n", __func__, params.fn_train_data);
}
printf("%s: number of training tokens: %d\n", __func__, (int) train_tokens.size());
struct my_llama_model model;
model.hparams.n_vocab = llama_n_vocab(lctx);
model.hparams.n_ctx = params.n_ctx;
model.hparams.n_embd = params.n_embd;
model.hparams.n_head = params.n_head;
model.hparams.n_layer = params.n_layer;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
model.hparams.n_ff = params.n_ff;
// llama.cpp requires n_rot to be exactly n_embd / n_head
model.hparams.n_rot = model.hparams.n_embd / model.hparams.n_head;
model.hparams.f_norm_rms_eps = params.f_norm_rms_eps;
model.hparams.rope_freq_base = params.rope_freq_base;
model.hparams.rope_freq_scale = params.rope_freq_scale;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
print_params(&model.hparams);
std::vector<size_t> token_noccurs;
std::vector<bool> token_notavail;
token_noccurs.resize(model.hparams.n_vocab, 0);
token_notavail.resize(model.hparams.n_vocab, true);
for (int i = 0; i < (int) train_tokens.size(); ++i) {
++token_noccurs[train_tokens[i]];
token_notavail[train_tokens[i]] = false;
}
std::vector<float> token_freq;
token_freq.resize(model.hparams.n_vocab, 0);
int n_unique_tokens = 0;
for (int i = 0; i < (int) token_noccurs.size(); ++i) {
token_freq[i] = (float) token_noccurs[i] / (float) train_tokens.size();
n_unique_tokens += (token_noccurs[i] > 0) ? 1 : 0;
}
printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens);
struct ggml_init_params lcparams;
lcparams.mem_size = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
lcparams.mem_buffer = NULL;
lcparams.no_alloc = false;
model.ctx = ggml_init(lcparams);
int n_tokens = model.hparams.n_ctx;
int n_vocab = model.hparams.n_vocab;
int n_batch = params.n_batch;
struct ggml_opt_context * opt = (struct ggml_opt_context *) alloca(sizeof(struct ggml_opt_context));
memset(opt, 0, sizeof(struct ggml_opt_context));
struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
opt_params_adam.print_forward_graph = false;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
opt_params_adam.print_backward_graph = false;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
opt_params_adam.n_threads = params.n_threads;
opt_params_adam.past = params.opt_past;
opt_params_adam.delta = params.opt_delta;
opt_params_adam.max_no_improvement = params.opt_max_no_improvement;
opt_params_adam.adam.n_iter = params.adam_n_iter;
opt_params_adam.adam.sched = 1.0f;
opt_params_adam.adam.alpha = params.adam_alpha;
opt_params_adam.adam.decay = params.adam_decay;
opt_params_adam.adam.decay_min_ndim = params.adam_decay_min_ndim;
opt_params_adam.adam.beta1 = params.adam_beta1;
opt_params_adam.adam.beta2 = params.adam_beta2;
opt_params_adam.adam.gclip = params.adam_gclip;
opt_params_adam.adam.eps_f = params.adam_eps_f;
opt_params_lbfgs.print_forward_graph = false;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
opt_params_lbfgs.print_backward_graph = false;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
opt_params_lbfgs.n_threads = params.n_threads;
opt_params_adam.past = params.opt_past;
opt_params_adam.delta = params.opt_delta;
opt_params_adam.max_no_improvement = params.opt_max_no_improvement;
opt_params_lbfgs.lbfgs.n_iter = params.lbfgs_n_iter;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
opt->ctx = model.ctx;
opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs;
printf("%s: init model\n", __func__);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
bool existed = load_checkpoint_file(params.fn_checkpoint_in, &model, opt);
if (!existed) {
init_model(&model);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
set_param_model(&model);
opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs;
opt->iter = model.train_its;
printf("%s: opt iter %d\n", __func__, opt->iter);
bool from_scratch = !existed;
if (from_scratch) {
randomize_model(&model, params.seed, 0.0f, 1.0f, -1.0f, +1.0f);
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
printf("used_mem model: %zu bytes\n", ggml_used_mem(model.ctx));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
// ggml_print_tensor_objects(model.ctx);
// TODO: use std::vector<uint8_t> intead of "new"
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
size_t compute_size = 1024ll*1024ll*1024ll*((size_t) params.mem_compute_gb);
uint8_t * compute_addr = new uint8_t[compute_size];
size_t size_buf_0 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute0_gb);
uint8_t * compute_buf_0 = new uint8_t[size_buf_0];
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_allocr * alloc = NULL;
if (params.use_alloc) {
static const size_t tensor_alignment = 32;
alloc = ggml_allocr_new(compute_buf_0, size_buf_0, tensor_alignment);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
GGML_ASSERT(n_tokens < (int) train_tokens.size());
std::vector<int> train_samples;
train_samples.push_back(0);
for (int i = 1; i < (int) train_tokens.size() - n_tokens; ++i) {
gguf : new file format with flexible meta data (beta) (#2398) * gguf : first API pass * gguf : read header + meta data * gguf : read tensor info * gguf : initial model loading - not tested * gguf : add gguf_get_tensor_name() * gguf : do not support passing existing ggml_context to gguf_init * gguf : simplify gguf_get_val * gguf : gguf.c is now part of ggml.c * gguf : read / write sample models * gguf : add comments * refactor : reduce code duplication and better API (#2415) * gguf : expose the gguf_type enum through the API for now * gguf : add array support * gguf.py : some code style changes * convert.py : start a new simplified implementation by removing old stuff * convert.py : remove GGML vocab + other obsolete stuff * GGUF : write tensor (#2426) * WIP: Write tensor * GGUF : Support writing tensors in Python * refactor : rm unused import and upd todos * fix : fix errors upd writing example * rm example.gguf * gitignore *.gguf * undo formatting * gguf : add gguf_find_key (#2438) * gguf.cpp : find key example * ggml.h : add gguf_find_key * ggml.c : add gguf_find_key * gguf : fix writing tensors * gguf : do not hardcode tensor names to read * gguf : write sample tensors to read * gguf : add tokenization constants * quick and dirty conversion example * gguf : fix writing gguf arrays * gguf : write tensors one by one and code reuse * gguf : fix writing gguf arrays * gguf : write tensors one by one * gguf : write tensors one by one * gguf : write tokenizer data * gguf : upd gguf conversion script * Update convert-llama-h5-to-gguf.py * gguf : handle already encoded string * ggml.h : get array str and f32 * ggml.c : get arr str and f32 * gguf.py : support any type * Update convert-llama-h5-to-gguf.py * gguf : fix set is not subscriptable * gguf : update convert-llama-h5-to-gguf.py * constants.py : add layer norm eps * gguf.py : add layer norm eps and merges * ggml.h : increase GGML_MAX_NAME to 64 * ggml.c : add gguf_get_arr_n * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Makefile : add gptneox gguf example * Update convert-llama-h5-to-gguf.py * add gptneox gguf example * Update convert-llama-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-gptneox-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * gguf : support custom alignment value * gguf : fix typo in function call * gguf : mmap tensor data example * fix : update convert-llama-h5-to-gguf.py * Update convert-llama-h5-to-gguf.py * convert-gptneox-h5-to-gguf.py : Special tokens * gptneox-main.cpp : special tokens * Update gptneox-main.cpp * constants.py : special tokens * gguf.py : accumulate kv and tensor info data + special tokens * convert-gptneox-h5-to-gguf.py : accumulate kv and ti + special tokens * gguf : gguf counterpart of llama-util.h * gguf-util.h : update note * convert-llama-h5-to-gguf.py : accumulate kv / ti + special tokens * convert-llama-h5-to-gguf.py : special tokens * Delete gptneox-common.cpp * Delete gptneox-common.h * convert-gptneox-h5-to-gguf.py : gpt2bpe tokenizer * gptneox-main.cpp : gpt2 bpe tokenizer * gpt2 bpe tokenizer (handles merges and unicode) * Makefile : remove gptneox-common * gguf.py : bytesarray for gpt2bpe tokenizer * cmpnct_gpt2bpe.hpp : comments * gguf.py : use custom alignment if present * gguf : minor stuff * Update gptneox-main.cpp * map tensor names * convert-gptneox-h5-to-gguf.py : map tensor names * convert-llama-h5-to-gguf.py : map tensor names * gptneox-main.cpp : map tensor names * gguf : start implementing libllama in GGUF (WIP) * gguf : start implementing libllama in GGUF (WIP) * rm binary commited by mistake * upd .gitignore * gguf : calculate n_mult * gguf : inference with 7B model working (WIP) * gguf : rm deprecated function * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : start implementing gguf_file_saver (WIP) * gguf : add gguf_get_kv_type * gguf : add gguf_get_kv_type * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver (WIP) * gguf : write metadata in gguf_file_saver * gguf : rm references to old file formats * gguf : shorter name for member variable * gguf : rm redundant method * gguf : get rid of n_mult, read n_ff from file * Update gguf_tensor_map.py * Update gptneox-main.cpp * gguf : rm references to old file magics * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : start implementing quantization (WIP) * gguf : quantization is working * gguf : roper closing of file * gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : no need to convert tensors twice * convert-llama-h5-to-gguf.py : no need to convert tensors twice * convert-gptneox-h5-to-gguf.py : simplify nbytes * convert-llama-h5-to-gguf.py : simplify nbytes * gptneox-main.cpp : n_layer --> n_block * constants.py : n_layer --> n_block * gguf.py : n_layer --> n_block * convert-gptneox-h5-to-gguf.py : n_layer --> n_block * convert-llama-h5-to-gguf.py : n_layer --> n_block * gptneox-main.cpp : n_layer --> n_block * Update gguf_tensor_map.py * convert-gptneox-h5-to-gguf.py : load model in parts to save memory * convert-llama-h5-to-gguf.py : load model in parts to save memory * convert : write more metadata for LLaMA * convert : rm quantization version * convert-gptneox-h5-to-gguf.py : add file_type key * gptneox-main.cpp : add file_type key * fix conflicts * gguf : add todos and comments * convert-gptneox-h5-to-gguf.py : tensor name map changes * Create gguf_namemap.py : tensor name map changes * Delete gguf_tensor_map.py * gptneox-main.cpp : tensor name map changes * convert-llama-h5-to-gguf.py : fixes * gguf.py : dont add empty strings * simple : minor style changes * gguf : use UNIX line ending * Create convert-llama-7b-pth-to-gguf.py * llama : sync gguf-llama.cpp with latest llama.cpp (#2608) * llama : sync gguf-llama.cpp with latest llama.cpp * minor : indentation + assert * llama : refactor gguf_buffer and gguf_ctx_buffer * llama : minor * gitignore : add gptneox-main * llama : tokenizer fixes (#2549) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * convert : update convert-new.py with tokenizer fixes (#2614) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * llama : sync gguf-llama with llama (#2613) * llama : sync gguf-llama with llama * tests : fix build + warnings (test-tokenizer-1 still fails) * tests : fix wstring_convert * convert : fix layer names * llama : sync gguf-llama.cpp * convert : update HF converter to new tokenizer voodoo magics * llama : update tokenizer style * convert-llama-h5-to-gguf.py : add token types * constants.py : add token types * gguf.py : add token types * convert-llama-7b-pth-to-gguf.py : add token types * gguf-llama.cpp : fix n_head_kv * convert-llama-h5-to-gguf.py : add 70b gqa support * gguf.py : add tensor data layout * convert-llama-h5-to-gguf.py : add tensor data layout * convert-llama-7b-pth-to-gguf.py : add tensor data layout * gptneox-main.cpp : add tensor data layout * convert-llama-h5-to-gguf.py : clarify the reverse permute * llama : refactor model loading code (#2620) * llama : style formatting + remove helper methods * llama : fix quantization using gguf tool * llama : simplify gguf_file_saver * llama : fix method names * llama : simplify write_header() * llama : no need to pass full file loader to the file saver just gguf_ctx * llama : gguf_file_saver write I32 * llama : refactor tensor names (#2622) * gguf: update tensor names searched in quantization * gguf : define tensor names as constants * gguf : initial write API (not tested yet) * gguf : write to file API (not tested) * gguf : initial write API ready + example * gguf : fix header write * gguf : fixes + simplify example + add ggml_nbytes_pad() * gguf : minor * llama : replace gguf_file_saver with new gguf write API * gguf : streaming support when writing files * gguf : remove oboslete write methods * gguf : remove obosolete gguf_get_arr_xxx API * llama : simplify gguf_file_loader * llama : move hparams and vocab from gguf_file_loader to llama_model_loader * llama : merge gguf-util.h in llama.cpp * llama : reorder definitions in .cpp to match .h * llama : minor simplifications * llama : refactor llama_model_loader (WIP) wip : remove ggml_ctx from llama_model_loader wip : merge gguf_file_loader in llama_model_loader * llama : fix shape prints * llama : fix Windows build + fix norm_rms_eps key * llama : throw error on missing KV paris in model meta data * llama : improve printing + log meta data * llama : switch print order of meta data --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> * gguf : deduplicate (#2629) * gguf : better type names * dedup : CPU + Metal is working * ggml : fix warnings about unused results * llama.cpp : fix line feed and compiler warning * llama : fix strncpy warning + note token_to_str does not write null * llama : restore the original load/save session implementation Will migrate this to GGUF in the future * convert-llama-h5-to-gguf.py : support alt ctx param name * ggml : assert when using ggml_mul with non-F32 src1 * examples : dedup simple --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> * gguf.py : merge all files in gguf.py * convert-new.py : pick #2427 for HF 70B support * examples/gguf : no need to keep q option for quantization any more * llama.cpp : print actual model size * llama.cpp : use ggml_elements() * convert-new.py : output gguf (#2635) * convert-new.py : output gguf (WIP) * convert-new.py : add gguf key-value pairs * llama : add hparams.ctx_train + no longer print ftype * convert-new.py : minor fixes * convert-new.py : vocab-only option should work now * llama : fix tokenizer to use llama_char_to_byte * tests : add new ggml-vocab-llama.gguf * convert-new.py : tensor name mapping * convert-new.py : add map for skipping tensor serialization * convert-new.py : convert script now works * gguf.py : pick some of the refactoring from #2644 * convert-new.py : minor fixes * convert.py : update to support GGUF output * Revert "ci : disable CI temporary to not waste energy" This reverts commit 7e82d25f40386540c2c15226300ad998ecd871ea. * convert.py : n_head_kv optional and .gguf file extension * convert.py : better always have n_head_kv and default it to n_head * llama : sync with recent PRs on master * editorconfig : ignore models folder ggml-ci * ci : update ".bin" to ".gguf" extension ggml-ci * llama : fix llama_model_loader memory leak * gptneox : move as a WIP example * llama : fix lambda capture ggml-ci * ggml : fix bug in gguf_set_kv ggml-ci * common.h : .bin --> .gguf * quantize-stats.cpp : .bin --> .gguf * convert.py : fix HF tensor permuting / unpacking ggml-ci * llama.cpp : typo * llama : throw error if gguf fails to init from file ggml-ci * llama : fix tensor name grepping during quantization ggml-ci * gguf.py : write tensors in a single pass (#2644) * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : single pass for writing tensors + refactoring writer * gguf : style fixes in simple conversion script * gguf : refactor gptneox conversion script * gguf : rename h5 to hf (for HuggingFace) * gguf : refactor pth to gguf conversion script * gguf : rm file_type key and method * gguf.py : fix vertical alignment * gguf.py : indentation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * convert-gptneox-hf-to-gguf.py : fixes * gguf.py : gptneox mapping * convert-llama-hf-to-gguf.py : fixes * convert-llama-7b-pth-to-gguf.py : fixes * ggml.h : reverse GGUF_MAGIC * gguf.py : reverse GGUF_MAGIC * test-tokenizer-0.cpp : fix warning * llama.cpp : print kv general.name * llama.cpp : get special token kv and linefeed token id * llama : print number of tensors per type + print arch + style * tests : update vocab file with new magic * editorconfig : fix whitespaces * llama : re-order functions * llama : remove C++ API + reorganize common source in /common dir * llama : minor API updates * llama : avoid hardcoded special tokens * llama : fix MPI build ggml-ci * llama : introduce enum llama_vocab_type + remove hardcoded string constants * convert-falcon-hf-to-gguf.py : falcon HF --> gguf conversion, not tested * falcon-main.cpp : falcon inference example * convert-falcon-hf-to-gguf.py : remove extra kv * convert-gptneox-hf-to-gguf.py : remove extra kv * convert-llama-7b-pth-to-gguf.py : remove extra kv * convert-llama-hf-to-gguf.py : remove extra kv * gguf.py : fix for falcon 40b * falcon-main.cpp : fix for falcon 40b * convert-falcon-hf-to-gguf.py : update ref * convert-falcon-hf-to-gguf.py : add tensor data layout * cmpnct_gpt2bpe.hpp : fixes * falcon-main.cpp : fixes * gptneox-main.cpp : fixes * cmpnct_gpt2bpe.hpp : remove non-general stuff * Update examples/server/README.md Co-authored-by: slaren <slarengh@gmail.com> * cmpnct_gpt2bpe.hpp : cleanup * convert-llama-hf-to-gguf.py : special tokens * convert-llama-7b-pth-to-gguf.py : special tokens * convert-permute-debug.py : permute debug print * convert-permute-debug-master.py : permute debug for master * convert-permute-debug.py : change permute type of attn_q * convert.py : 70b model working (change attn_q permute) * Delete convert-permute-debug-master.py * Delete convert-permute-debug.py * convert-llama-hf-to-gguf.py : fix attn_q permute * gguf.py : fix rope scale kv * convert-llama-hf-to-gguf.py : rope scale and added tokens * convert-llama-7b-pth-to-gguf.py : rope scale and added tokens * llama.cpp : use rope scale kv * convert-llama-7b-pth-to-gguf.py : rope scale fix * convert-llama-hf-to-gguf.py : rope scale fix * py : fix whitespace * gguf : add Python script to convert GGMLv3 LLaMA models to GGUF (#2682) * First pass at converting GGMLv3 LLaMA models to GGUF * Cleanups, better output during conversion * Fix vocab space conversion logic * More vocab conversion fixes * Add description to converted GGUF files * Improve help text, expand warning * Allow specifying name and description for output GGUF * Allow overriding vocab and hyperparams from original model metadata * Use correct params override var name * Fix wrong type size for Q8_K Better handling of original style metadata * Set default value for gguf add_tensor raw_shape KW arg * llama : improve token type support (#2668) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * llama : add API for token type ggml-ci * tests : use new tokenizer type API (#2692) * Merge tokenizer fixes into the gguf branch. * Add test vocabularies * Adapt convert-new.py (and fix a clang-cl compiler error on windows) * Improved tokenizer test But does it work on MacOS? * Improve token type support - Added @klosax code to convert.py - Improved token type support in vocabulary * Exclude platform dependent tests * More sentencepiece compatibility by eliminating magic numbers * Restored accidentally removed comment * Improve commentary * Use token type API in test-tokenizer-1.cpp * py : cosmetics * readme : add notice about new file format ggml-ci --------- Co-authored-by: M. Yusuf Sarıgöz <yusufsarigoz@gmail.com> Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: goerch <jhr.walter@t-online.de> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
2023-08-21 22:07:43 +02:00
if (!params.samples_start_after_nl || (train_tokens[i-1] == llama_token_nl(lctx))) {
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train_samples.push_back(i);
}
}
shuffle_ints(train_samples.data(), train_samples.data() + train_samples.size());
for (int i = 0; i < (int) train_samples.size(); ++i) {
GGML_ASSERT(train_samples[i]+n_tokens-1 < (int) train_tokens.size());
}
printf("%s: begin training\n", __func__);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct opt_callback_data opt_cb_data;
opt_cb_data.params = &params;
opt_cb_data.opt = opt;
opt_cb_data.lctx = lctx;
opt_cb_data.tokens_data = train_tokens.data();
opt_cb_data.tokens_size = train_tokens.size();
opt_cb_data.samples_data = train_samples.data();
opt_cb_data.samples_size = train_samples.size();
opt_cb_data.shuffle_countdown = train_samples.size();
opt_cb_data.tokens_input = NULL;
opt_cb_data.target_logits = NULL;
opt_cb_data.target_probs = NULL;
int64_t t0 = ggml_time_ms();
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
for (int ex = 0; ex < params.n_examples; ++ex) {
if (ex*n_batch >= (int) train_samples.size()) {
shuffle_ints(train_samples.data(), train_samples.data() + train_samples.size());
for (int i = 0; i < (int) train_samples.size(); ++i) {
GGML_ASSERT(train_samples[i]+n_tokens-1 < (int) train_tokens.size());
}
}
struct ggml_init_params cparams = {
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
compute_size, // mem_size
compute_addr, // mem_buffer
false, // no_alloc
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
};
struct ggml_context * ctx0 = ggml_init(cparams);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_no_alloc(ctx0, false);
// don't use alloc for input tensors, so we can safely fill them with data
//struct ggml_tensor * after_opt_best_samples = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
//struct ggml_tensor * after_opt_probs = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch);
struct ggml_tensor * tokens_input = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch);
struct ggml_tensor * target_logits = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch);
struct ggml_tensor * target_probs = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_set_no_alloc(ctx0, (alloc != NULL));
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (alloc) {
ggml_allocr_reset(alloc);
}
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
opt_cb_data.tokens_input = tokens_input;
opt_cb_data.target_logits = target_logits;
opt_cb_data.target_probs = target_probs;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int n_past = 0;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gb = ggml_new_graph(ctx0);
struct ggml_cgraph * gb_tmp = params.use_checkpointing
? ggml_new_graph(ctx0)
: NULL;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
GGML_ASSERT(n_past == 0);
struct ggml_tensor * loss = NULL;
struct ggml_tensor * logits = NULL;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
loss = llama_build_train_graphs(
&model, alloc, ctx0,
gf, gb, gb_tmp,
&logits, tokens_input, target_probs,
n_tokens, n_batch,
params.use_flash,
params.use_checkpointing
);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
size_t used_mem_before_opt = ggml_used_mem(ctx0);
opt->params.adam.sched = (opt->iter < params.warmup)
? (float) opt->iter / (float) params.warmup
: cosine_decay_restart(
params.cos_decay_steps,
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.cos_decay_min,
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
opt->iter - params.warmup,
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
params.cos_decay_restart,
params.enable_restart);
float min_sched = params.adam_min_alpha / params.adam_alpha;
opt->params.adam.sched = min_sched + opt->params.adam.sched * (1.0f - min_sched);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
printf("%s: opt->params.adam.sched %.5f\n", __func__, opt->params.adam.sched);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_opt_resume_g(ctx0, opt, loss, gf, gb, &opt_callback, (void *) &opt_cb_data);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
size_t used_mem_after_opt = ggml_used_mem(ctx0);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int n_iter = params.use_adam ? params.adam_n_iter : params.lbfgs_n_iter;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
model.train_its = opt->iter;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
model.train_samples += n_batch * n_iter;
model.train_tokens += n_batch * n_tokens * n_iter;
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (params.print_info_interval > 0 && ex % params.print_info_interval == 0) {
printf("Example %d, opt iter %d\n", ex, opt->iter);
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
printf("error_before_opt: %.6f\n", opt->loss_before);
printf("error_after_opt: %.6f\n", opt->loss_after);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
printf("used_mem_before_opt: %zu bytes\n", used_mem_before_opt);
printf("used_mem_after_opt: %zu bytes\n", used_mem_after_opt);
}
ggml_free(ctx0);
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
int64_t t1 = ggml_time_ms();
int64_t d = t1-t0;
double dd = (double) d * 1e-3;
printf("%s: total training time=%f seconds\n", __func__, dd);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
if (params.n_examples > 0) {
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
save_checkpoint_file(params.fn_checkpoint_out, params.fn_vocab_model, &model, opt);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
if (strlen(params.fn_model_out) > 0) {
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
save_llama_model_file(params.fn_model_out, params.fn_vocab_model, &model);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
if (alloc) {
ggml_allocr_free(alloc);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
}
delete[] compute_addr;
delete[] compute_buf_0;
train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-08-28 21:51:47 +02:00
ggml_free(model.ctx);
llama_free(lctx);
llama_free_model(lmodel);
train : improved training-from-scratch example (#1652) * add python wrapper https://gist.github.com/abetlen/2b90e5f153f6efd00931d098de5c73ce * fix decoding error. adds errors=ignore parameter * add python bindings for functions to get and set the whole llama state (rng, logits, embedding and kv_cache) * update python bindings * add text generating baby-llama from scratch example * fix race condition bug in ggml_compute_forward_diag_mask_f32 * implement ggml_soft_max_back for more performant backward pass of soft_max avoids creating big intermediate matrices of size n_embd x n_embd for llama layers and n_vocab x n_vocab for cross entropy loss * improve softmax backward pass go from quadratic runtime to linear runtime by simplifying the formulas * fix race condition bug in non-inplace ggml_compute_forward_diag_mask_f32 memcpy needs to be synchronized across threads to avoid race conditions. => do it in INIT phase * fix bug in ggml_compute_forward_soft_max_back_f32 on DEBUG build * improve performance of mul_mat backward pass avoid transpose by using mul_mat with swapped arguments * avoid printing too much newlines in baby-llama-text * activate threading in baby-llama-text * add ggml_out_prod and use it for mul_mat backward pass for improved performance performance stats report improvement from 37 seconds to 16 seconds runtime during my training tests * better weight initialization improves training convergence at start * better weight initialization improves training convergence at start * improve ggml_out_prod performance - change iteration order (>15s -> 10s runtime) - parallelize over one more dimension: over dst matrix rows (10s -> <5s runtime) * add llama sampler, shuffle samples and constrain sampling to tokens occurring in train data * fix get_samples call, add model tensor names, increase model size, start training samples after newline * save train trained model to checkpoint and load model to be trained from checkpoint * use inplace functions where possible * initialize rng with srand * use different arguments for input and output checkpoint * ggml fixes to support backward pass on inplace operations * remove duplicate include * fix cross entropy loss - add target probabilities for each sample which is then used in cross entropy loss * print used memory before and after optimization * sample with non-greedy sampling parameters at the end of training * add cmake target for baby-llama-text * add ggml_add1_inplace to header * enable gradient propagation for inplace add1 and scale operations those functions backward passes don't need the original src0, so they also work when forward is inplace * implement AdamW in ggml_opt_adam by adding weight decay parameter (default 0.001f) also add a schedule parameter (default 1.0f) that can be used to scale alpha and decay according to learning schedule. setting the decay parameter to zero disables AdamW resulting in normal Adam optimizer. since the difference between Adam and AdamW is minimal it is not implemented as another optimizer, but integrated into the existing Adam optimizer. * use inplace operations in cross_entropy_loss * fix random weight initialization scale * add missing default parameters for adam optimizer * add ggml_opt_context, so that we can properly resume training otherwise the optimizer states, tracking statistics about the error function and its derivates, will reset to zero each time ggml_opt is called, hindering convergence on resumed training. now the optimizer context and all its memory is stored in a separate struct. * fix bug in llama_sample_token_mirostat_v2 when all candidates are filtered out through mu threshold, the following soft_max operation will fail. so keep at least one. * add forward function without using cache, for more performant training during training on whole samples no cache is required. removing the cache and simplifying the remaining code results in performance and memory usage improvement. * print suppressed newline tokens as string "\n" printing too much actual newlines is suppressed to avoid flooding the console. * store optimizer state in training checkpoint and add learning schedule persistent optimizer state allows to resume training without resetting the optimizer learning schedule consists of linear warmup ramp followed by cosine decay with restarts * remove unused functions * fix bug in get_samples which corrupted training targets * save checkpoint only when it was trained * simplify code * remove trailing whitespace * simplify backward pass for SQRT * replace inefficient repeat backward pass with dedicated repeat_back operation * add ggml_cross_entropy_loss with backward pass for faster training cross entropy loss can also be implemented using softmax and log, but as dedicated operation it is faster and especially avoids unnecessary memory overhead. * add tests for cross_entropy_loss backward pass finite differences regularly results in estimated gradient of zero, despite the backward pass giving non zero gradient. _probably_ the finite differences fails due to numerical issues * use ggml_cross_entropy_loss in text training example * remove trailing whitespace * slightly improve how cross entropy loss is compute btw: directly implemented cross entropy loss seems to have way lower magnitudes than when implemented with softmax and log. probably the input to log gets closer to zero due to float numerics. maybe the multiplication by (1.0-eps)/sum is more accurate.. * add llama_get_vocab to get the vocabulary as output parameters * set default model.type for unknown models with few layers * add export of training checkpoint to llama compatible model file * get vocabulary for exporting training checkpoint to llama compatible model file * implement backward pass of flash attention * bugfixes for backward pass of flash attention * test flash attention backward pass need to set loose error bounds to pass. the finitie differences are close to numeric limits and often return quite different values than the backward pass. reducing eps further lets the gradients vanish completely. likewise setting eps to big results in wronger values. the softmax in the middle of the function is probably the most responsible for the numeric issues using finite differences. * add option to train with flash attention and move options to the top of the main function training from scratch also works with flash attention training convergence and generation results after fix number of iterations are worse than when not using flash attention. maybe there still lingers a bug in the flash attention backward pass? but training works, just with slower convergence. flash attention is still worth to use, because it requires way less memory and is faster with high n_ctx * add train_params and command line option parser * remove unnecessary comments * add train params to specify memory size * remove python bindings * rename baby-llama-text to train-text-from-scratch * replace auto parameters in lambda function * add #include <climits> * add explicit cast to fix compile error "error: non-constant-expression cannot be narrowed from type 'int64_t' (aka 'long long') to 'uint32_t' (aka 'unsigned int') in initializer list [-Wc++11-narrowing]" * remove trailing whitespace * add ggml_opt_resume_g which accepts forward and backward cgraphs * fix formulas in comments * bug fix for ggml_compute_forward_get_rows_back_f32 the result should be set to zero, not to whatever data is in opt0 * improve training memory usage with scratch buffers instead of relying on the automatic backward pass, we manually create the graph for the backward pass. it turns out that all backward pass operations need only temporary memory which can be reused after each layer. will compute backward pass for ALL model parameters * add option to use scratch buffers in training or not make it configurable because currently training with scratch buffers implies flash attention and optimization over all parameters. * ci : disable temporary * store view offset and permute axes in opt[0] instead of storing it in padding use memcpy to store offset, because offset is of type size_t. when storing it as int32_t offset would have to be smaller than 2^31 which is not necessarily true. * minor : fix compile warnings + minor style changes * fix bug in threaded indices calculation of ggml_compute_forward_flash_attn_back_f32 * store view offset like in master branch * bug fix in forward_batch_wo_cache_flash_attn_train * scratch buffer bug fixes in forward_batch_wo_cache_flash_attn_train data of permute and reshape is the same as their input. if we want to preserve the output of permute/reshape, we also need to preserve their inputs. replace reshape(src0, src1) with reshape_nd calls so that we don't need src1. replace (temporary) t03 with ggml_repeat(ctx0, layer.attention_norm, t02). in the future we could also use the new broadcasting ggml_mul to avoid these repeat calls. for this we need backward pass of broadcasting ggml_mul. * remove unnecessary scratch buffer 0 buf 0 is persistent memory, so we can just disable scratch for this by using buf -1 * avoid creating unnecessary grad tensors previously we need to create grads for model parameters, so that expand(..) correctly populates cgraph->leafs & cgraph->grads this wasted memory, because unnecessary grad for each op were automatically created: the automatically generated grad was unnecessary because we later manually set the grad (e.g. t35->grad = expand(gb, ...) ). this discarded the automatically generated grad resulting in wasted memory. improved this by changing expand(..) to not use ggml_build_forward_expand. expand set cgraph->nodes but not the leafs. cgraph->leafs & cgraph->grads are set in another pass after the last expand call. * print used training seed * zero initialize gfbuf and gbbuf * ci : re-enable workflows + add README for training --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-13 21:04:40 +02:00
return 0;
}