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"
|
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>
|
|
|
|
|
|
2023-06-16 20:23:53 +02:00
|
|
|
#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
|
|
|
|
2023-08-21 22:07:43 +02:00
|
|
|
static const float rms_norm_eps = 1e-5f;
|
2023-07-24 17:57:12 +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
|
|
|
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);
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-07 18:24:01 +02:00
|
|
|
void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
|
|
|
|
|
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
|
|
|
|
|
|
|
|
|
|
if (plan.work_size > 0) {
|
|
|
|
|
buf.resize(plan.work_size);
|
|
|
|
|
plan.work_data = buf.data();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ggml_graph_compute(graph, &plan);
|
|
|
|
|
}
|
|
|
|
|
|
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 * 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 llama_vocab {
|
|
|
|
|
using id = int32_t;
|
|
|
|
|
using token = std::string;
|
2023-08-21 22:07:43 +02:00
|
|
|
using ttype = llama_token_type;
|
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
|
|
|
|
2023-08-21 22:07:43 +02:00
|
|
|
struct token_data {
|
|
|
|
|
token text;
|
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 score;
|
2023-08-21 22:07:43 +02:00
|
|
|
ttype type;
|
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
|
|
|
};
|
|
|
|
|
|
|
|
|
|
std::unordered_map<token, id> token_to_id;
|
2023-08-21 22:07:43 +02:00
|
|
|
std::vector<token_data> id_to_token;
|
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_hparams {
|
|
|
|
|
uint32_t n_vocab = 32000;
|
|
|
|
|
uint32_t n_ctx = 512; // this is provided as user input?
|
|
|
|
|
uint32_t n_embd = 4096;
|
|
|
|
|
uint32_t n_mult = 4;
|
|
|
|
|
uint32_t n_head = 32;
|
|
|
|
|
uint32_t n_layer = 32;
|
|
|
|
|
uint32_t n_rot = 64;
|
|
|
|
|
|
|
|
|
|
bool operator!=(const my_llama_hparams& other) const {
|
|
|
|
|
return memcmp(this, &other, sizeof(my_llama_hparams));
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
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_kv_cache {
|
|
|
|
|
struct ggml_context * ctx = NULL;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * k;
|
|
|
|
|
struct ggml_tensor * v;
|
|
|
|
|
|
|
|
|
|
// llama_ctx_buffer buf;
|
|
|
|
|
|
|
|
|
|
int n; // number of tokens currently in the cache
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
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;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
|
|
|
|
|
const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
|
|
|
|
|
return n_ff;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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_mult: %d\n", __func__, params->n_mult);
|
|
|
|
|
printf("%s: n_head: %d\n", __func__, params->n_head);
|
|
|
|
|
printf("%s: n_ff: %d\n", __func__, get_n_ff(params));
|
|
|
|
|
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;
|
|
|
|
|
|
|
|
|
|
const uint32_t n_ff = get_n_ff(&hparams);
|
|
|
|
|
|
|
|
|
|
struct ggml_context * ctx = model->ctx;
|
|
|
|
|
|
|
|
|
|
model->train_its = 0;
|
|
|
|
|
model->train_samples = 0;
|
|
|
|
|
model->train_tokens = 0;
|
|
|
|
|
|
|
|
|
|
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);
|
|
|
|
|
|
|
|
|
|
ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
|
|
|
|
|
ggml_set_name(model->norm, "norm.weight");
|
|
|
|
|
ggml_set_name(model->output, "output.weight");
|
|
|
|
|
|
|
|
|
|
model->layers.resize(n_layer);
|
|
|
|
|
for (uint32_t i = 0; i < n_layer; ++i) {
|
|
|
|
|
auto & layer = model->layers[i];
|
|
|
|
|
|
|
|
|
|
std::string layers_i = "layers." + std::to_string(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);
|
|
|
|
|
|
|
|
|
|
ggml_set_name(layer.attention_norm, (layers_i + ".attention_norm.weight").c_str());
|
|
|
|
|
|
|
|
|
|
ggml_set_name(layer.wq, (layers_i + ".attention.wq.weight").c_str());
|
|
|
|
|
ggml_set_name(layer.wk, (layers_i + ".attention.wk.weight").c_str());
|
|
|
|
|
ggml_set_name(layer.wv, (layers_i + ".attention.wv.weight").c_str());
|
|
|
|
|
ggml_set_name(layer.wo, (layers_i + ".attention.wo.weight").c_str());
|
|
|
|
|
|
|
|
|
|
ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str());
|
|
|
|
|
|
2023-06-26 21:45:32 +02:00
|
|
|
ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str());
|
|
|
|
|
ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
|
|
|
|
|
ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.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
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
bool init_kv_cache(struct my_llama_kv_cache* cache, struct my_llama_model * model, int n_batch) {
|
|
|
|
|
const auto & hparams = model->hparams;
|
|
|
|
|
|
|
|
|
|
const uint32_t n_ctx = hparams.n_ctx;
|
|
|
|
|
const uint32_t n_embd = hparams.n_embd;
|
|
|
|
|
const uint32_t n_layer = hparams.n_layer;
|
|
|
|
|
|
|
|
|
|
const int64_t n_mem = n_layer*n_ctx*n_batch;
|
|
|
|
|
const int64_t n_elements = n_embd*n_mem;
|
|
|
|
|
|
|
|
|
|
// cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
|
|
|
|
|
|
|
|
|
|
// struct ggml_init_params params;
|
|
|
|
|
// params.mem_size = cache.buf.size;
|
|
|
|
|
// params.mem_buffer = cache.buf.addr;
|
|
|
|
|
// params.no_alloc = false;
|
|
|
|
|
if (!cache->ctx) {
|
|
|
|
|
struct ggml_init_params params;
|
|
|
|
|
params.mem_size = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
|
|
|
|
|
params.mem_buffer = NULL;
|
|
|
|
|
params.no_alloc = false;
|
|
|
|
|
|
|
|
|
|
cache->ctx = ggml_init(params);
|
|
|
|
|
|
|
|
|
|
if (!cache->ctx) {
|
|
|
|
|
fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
|
|
|
|
|
cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
|
|
|
|
|
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * forward(
|
|
|
|
|
struct my_llama_model * model,
|
|
|
|
|
struct my_llama_kv_cache * cache,
|
|
|
|
|
struct ggml_context * ctx0,
|
|
|
|
|
struct ggml_cgraph * gf,
|
|
|
|
|
struct ggml_tensor * tokens_input,
|
|
|
|
|
const int n_tokens,
|
|
|
|
|
const int n_past) {
|
|
|
|
|
|
|
|
|
|
const int N = n_tokens;
|
|
|
|
|
|
|
|
|
|
struct my_llama_kv_cache& kv_self = *cache;
|
|
|
|
|
const auto & hparams = model->hparams;
|
|
|
|
|
const int n_ctx = hparams.n_ctx;
|
|
|
|
|
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;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
|
|
|
|
memcpy(tokens->data, tokens_input->data, N*ggml_element_size(tokens));
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * kc = kv_self.k;
|
|
|
|
|
struct ggml_tensor * vc = kv_self.v;
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N,1,1]
|
|
|
|
|
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cur;
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 0);
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
|
|
|
|
|
// cur = attention_norm*cur
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// self-attention
|
|
|
|
|
{
|
|
|
|
|
// compute Q and K and RoPE them
|
|
|
|
|
// wq shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// wk shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, 1]
|
|
|
|
|
// Kcur shape [n_embd/n_head, n_head, N, 1]
|
2023-06-26 23:37:13 +02:00
|
|
|
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
|
|
|
|
|
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 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
|
|
|
|
|
|
|
|
// store key and value to memory
|
|
|
|
|
{
|
|
|
|
|
// compute the transposed [N, n_embd] V matrix
|
|
|
|
|
// wv shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// Vcur shape [n_embd, N, 1, 1]
|
|
|
|
|
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wv, cur), n_embd, N)));
|
|
|
|
|
|
|
|
|
|
// kv_self.k shape [n_embd * n_ctx * n_layer, 1]
|
|
|
|
|
// kv_self.v shape [n_embd * n_ctx * n_layer, 1]
|
|
|
|
|
// k shape [n_embd * N, 1] == kv_self.k[:,n_past:n_past+N,il,0]
|
|
|
|
|
// v shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0]
|
|
|
|
|
|
|
|
|
|
/* {
|
|
|
|
|
struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
|
|
|
|
|
struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
|
|
|
|
|
( n_ctx)*ggml_element_size(kv_self.v),
|
|
|
|
|
(il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
|
|
|
|
|
|
|
|
|
|
// important: storing RoPE-ed version of K in the KV cache!
|
|
|
|
|
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
|
|
|
|
|
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
|
|
|
|
|
} //*/
|
|
|
|
|
|
|
|
|
|
kc = ggml_set_1d_inplace(ctx0, kc, ggml_reshape_1d(ctx0, Kcur, n_embd*N), (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
|
|
|
|
|
vc = ggml_set_2d_inplace(ctx0, vc, Vcur, ( n_ctx)*ggml_element_size(kv_self.v),
|
|
|
|
|
(il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, 1]
|
|
|
|
|
// Q shape [n_embd/n_head, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * Q =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Qcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
|
|
|
|
|
// kv_self.k shape [n_embd * n_ctx * n_layer, 1]
|
|
|
|
|
// K shape [n_embd/n_head, n_past + N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * K =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
ggml_reshape_3d(ctx0,
|
|
|
|
|
ggml_view_1d(ctx0, kc, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kc)*n_embd),
|
|
|
|
|
n_embd/n_head, n_head, n_past + N),
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
|
|
|
|
|
// K * Q
|
|
|
|
|
// KQ shape [n_past + N, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
|
|
|
|
|
|
|
|
|
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
|
|
|
|
// KQ_scaled shape [n_past + N, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * KQ_scaled =
|
|
|
|
|
ggml_scale(ctx0,
|
|
|
|
|
KQ,
|
|
|
|
|
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
|
|
|
|
|
|
|
|
|
|
// KQ_masked = mask_past(KQ_scaled)
|
|
|
|
|
// KQ_masked shape [n_past + N, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
|
|
|
|
|
|
|
|
|
|
// KQ = soft_max(KQ_masked)
|
|
|
|
|
// KQ_soft_max shape [n_past + N, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
|
|
|
|
|
|
|
|
|
|
// split cached V into n_head heads
|
|
|
|
|
//// V shape [n_past + N, n_embd/n_head, n_head, 1]
|
|
|
|
|
// V shape [n_past + N, n_embd/n_head, n_head, 1] == kv_self.v[:,:(n_past+N),il,1]
|
|
|
|
|
struct ggml_tensor * V =
|
|
|
|
|
ggml_view_3d(ctx0, vc,
|
|
|
|
|
n_past + N, n_embd/n_head, n_head,
|
|
|
|
|
n_ctx*ggml_element_size(vc),
|
|
|
|
|
n_ctx*ggml_element_size(vc)*n_embd/n_head,
|
|
|
|
|
il*n_ctx*ggml_element_size(vc)*n_embd);
|
|
|
|
|
|
|
|
|
|
// KQV shape [n_embd/n_head, N, n_head, 1]
|
|
|
|
|
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
|
|
|
|
|
|
|
|
|
|
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
|
|
|
|
// KQV_merged shape [n_embd/n_head, n_head, N, 1]
|
|
|
|
|
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
|
|
|
|
// KQV_merged shape
|
|
|
|
|
|
|
|
|
|
// cur = KQV_merged.contiguous().view(n_embd, N)
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
|
|
|
|
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N);
|
|
|
|
|
// cur = ggml_cpy(ctx0,
|
|
|
|
|
// KQV_merged,
|
|
|
|
|
// ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
|
|
|
|
|
|
|
|
|
// projection (no bias)
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].wo,
|
|
|
|
|
cur);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 1);
|
|
|
|
|
|
|
|
|
|
// inpFF shape [n_embd,N,1,1]
|
|
|
|
|
struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
|
|
|
|
|
|
|
|
|
|
// feed-forward network
|
|
|
|
|
{
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_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
|
|
|
|
|
|
|
|
// cur = ffn_norm*cur
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// tmp shape [n_ff,N,1,1]
|
|
|
|
|
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w3,
|
|
|
|
|
cur);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w1,
|
|
|
|
|
cur);
|
|
|
|
|
|
|
|
|
|
// SILU activation
|
|
|
|
|
// cur shape [n_ff,N,1,1]
|
|
|
|
|
cur = ggml_silu(ctx0, cur);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0, cur, tmp);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w2,
|
|
|
|
|
cur);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N,1,1]
|
|
|
|
|
cur = ggml_add(ctx0, cur, inpFF);
|
|
|
|
|
|
|
|
|
|
// input for next layer
|
|
|
|
|
// inpL shape [n_embd,N,1,1]
|
|
|
|
|
inpL = cur;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
|
|
|
|
|
// inpL = norm*inpL
|
|
|
|
|
// inpL shape [n_embd,N,1,1]
|
|
|
|
|
inpL = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->norm, inpL),
|
|
|
|
|
inpL);
|
|
|
|
|
|
|
|
|
|
//embeddings = inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lm_head
|
|
|
|
|
// inpL shape [n_vocab,N,1,1]
|
|
|
|
|
inpL = ggml_mul_mat(ctx0, model->output, inpL);
|
|
|
|
|
|
|
|
|
|
// run the computation
|
|
|
|
|
ggml_build_forward_expand(gf, inpL);
|
|
|
|
|
|
|
|
|
|
return inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * forward_batch(
|
|
|
|
|
struct my_llama_model * model,
|
|
|
|
|
struct my_llama_kv_cache * cache,
|
|
|
|
|
struct ggml_context * ctx0,
|
|
|
|
|
struct ggml_cgraph * gf,
|
|
|
|
|
struct ggml_tensor * tokens_input,
|
|
|
|
|
const int n_tokens,
|
|
|
|
|
const int n_past,
|
|
|
|
|
const int n_batch) {
|
|
|
|
|
|
|
|
|
|
const int N = n_tokens;
|
|
|
|
|
|
|
|
|
|
struct my_llama_kv_cache& kv_self = *cache;
|
|
|
|
|
const auto & hparams = model->hparams;
|
|
|
|
|
const int n_ctx = hparams.n_ctx;
|
|
|
|
|
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;
|
|
|
|
|
const int n_ff = get_n_ff(&hparams);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
|
|
|
|
|
memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * kc = kv_self.k;
|
|
|
|
|
struct ggml_tensor * vc = kv_self.v;
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1]
|
|
|
|
|
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cur;
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 0);
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = attention_norm*cur
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// self-attention
|
|
|
|
|
{
|
|
|
|
|
// compute Q and K and RoPE them
|
|
|
|
|
// wq shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// wk shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
// Kcur shape [n_embd/n_head, n_head, N, n_batch]
|
2023-06-26 23:37:13 +02:00
|
|
|
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
|
|
|
|
|
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 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
|
|
|
assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
|
|
|
|
|
// store key and value to memory
|
|
|
|
|
{
|
|
|
|
|
// compute the transposed [N, n_embd] V matrix
|
|
|
|
|
// wv shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// Vcur shape [N, n_embd, n_batch, 1]
|
|
|
|
|
struct ggml_tensor * Vcur = ggml_cont(ctx0,
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
ggml_reshape_3d(ctx0,
|
|
|
|
|
ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].wv,
|
|
|
|
|
cur),
|
|
|
|
|
n_embd, N, n_batch),
|
|
|
|
|
1, 0, 2, 3));
|
|
|
|
|
assert_shape_3d(Vcur, N, n_embd, n_batch);
|
|
|
|
|
|
|
|
|
|
// kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
|
|
|
|
|
// kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
|
|
|
|
|
// k shape [n_embd * N, n_batch] == kv_self.k[:,n_past:n_past+N,:,il]
|
|
|
|
|
// v shape [N, n_embd, n_batch, 1] == kv_self.v[:,n_past:n_past+N,:,il]
|
|
|
|
|
|
|
|
|
|
/* {
|
|
|
|
|
struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
|
|
|
|
|
struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
|
|
|
|
|
( n_ctx)*ggml_element_size(kv_self.v),
|
|
|
|
|
(il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
|
|
|
|
|
|
|
|
|
|
// important: storing RoPE-ed version of K in the KV cache!
|
|
|
|
|
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
|
|
|
|
|
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
|
|
|
|
|
} //*/
|
|
|
|
|
|
|
|
|
|
kc = ggml_set_2d_inplace(ctx0, kc,
|
|
|
|
|
ggml_reshape_2d(ctx0, Kcur, n_embd*N, n_batch),
|
|
|
|
|
ggml_element_size(kc)*n_embd*n_ctx,
|
|
|
|
|
(ggml_element_size(kc)*n_embd)*(il*n_batch*n_ctx + n_past));
|
|
|
|
|
vc = ggml_set_2d_inplace(ctx0, vc,
|
|
|
|
|
ggml_reshape_2d(ctx0, Vcur, N*n_embd, n_batch),
|
|
|
|
|
ggml_element_size(vc)*n_ctx*n_embd,
|
|
|
|
|
ggml_element_size(vc)*(n_past + il*n_embd*n_batch*n_ctx));
|
|
|
|
|
|
|
|
|
|
assert_shape_1d(kc, n_embd * n_ctx * n_batch * n_layer);
|
|
|
|
|
assert_shape_1d(vc, n_embd * n_ctx * n_batch * n_layer);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
// Q shape [n_embd/n_head, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * Q =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Qcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
|
|
|
|
|
// K shape [n_embd/n_head, n_past + N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * K =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
ggml_reshape_4d(ctx0,
|
|
|
|
|
ggml_view_3d(ctx0,
|
|
|
|
|
kc,
|
|
|
|
|
n_embd,
|
|
|
|
|
(n_past + N),
|
|
|
|
|
n_batch,
|
|
|
|
|
n_embd*ggml_element_size(kc),
|
|
|
|
|
n_ctx*n_embd*ggml_element_size(kc),
|
|
|
|
|
il*n_batch*n_ctx*n_embd*ggml_element_size(kc)),
|
|
|
|
|
n_embd/n_head, n_head, n_past + N, n_batch),
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(K, n_embd/n_head, n_past + N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// K * Q
|
|
|
|
|
// KQ shape [n_past + N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
|
|
|
|
assert_shape_4d(KQ, n_past + N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
|
|
|
|
// KQ_scaled shape [n_past + N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_scaled =
|
|
|
|
|
ggml_scale_inplace(ctx0,
|
|
|
|
|
KQ,
|
|
|
|
|
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
|
|
|
|
|
assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ_masked = mask_past(KQ_scaled)
|
|
|
|
|
// KQ_masked shape [n_past + N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
|
|
|
|
assert_shape_4d(KQ_masked, n_past + N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ = soft_max(KQ_masked)
|
|
|
|
|
// KQ_soft_max shape [n_past + N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
|
|
|
|
|
assert_shape_4d(KQ_soft_max, n_past + N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// split cached V into n_head heads
|
|
|
|
|
// kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
|
|
|
|
|
// V shape [n_past + N, n_embd/n_head, n_head, n_batch] == kv_self.v[:(n_past+N),:,:,il]
|
|
|
|
|
struct ggml_tensor * V =
|
|
|
|
|
ggml_view_4d(ctx0, vc,
|
|
|
|
|
n_past + N, n_embd/n_head, n_head, n_batch,
|
|
|
|
|
ggml_element_size(vc)*n_ctx,
|
|
|
|
|
ggml_element_size(vc)*n_ctx*n_embd/n_head,
|
|
|
|
|
ggml_element_size(vc)*n_ctx*n_embd,
|
|
|
|
|
il*n_batch*n_ctx*n_embd*ggml_element_size(vc));
|
|
|
|
|
assert_shape_4d(V, n_past + N, n_embd/n_head, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQV shape [n_embd/n_head, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
|
|
|
|
|
assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
|
|
|
|
// KQV_merged shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
// KQV_merged shape
|
|
|
|
|
|
|
|
|
|
// cur = KQV_merged.contiguous().view(n_embd, N)
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
// cur = ggml_cpy(ctx0,
|
|
|
|
|
// KQV_merged,
|
|
|
|
|
// ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
|
|
|
|
|
|
|
|
|
// projection (no bias)
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].wo,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 1);
|
|
|
|
|
|
|
|
|
|
// inpFF shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
|
|
|
|
|
assert_shape_2d(inpFF, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// feed-forward network
|
|
|
|
|
{
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = ffn_norm*cur
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// tmp shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w3,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(tmp, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w1,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// SILU activation
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_silu(ctx0, cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0, cur, tmp);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w2,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_add_inplace(ctx0, cur, inpFF);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// input for next layer
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
inpL = cur;
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// inpL = norm*inpL
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
inpL = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->norm, inpL),
|
|
|
|
|
inpL);
|
|
|
|
|
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
//embeddings = inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lm_head
|
|
|
|
|
// inpL shape [n_vocab,N*n_batch,1,1]
|
|
|
|
|
inpL = ggml_mul_mat(ctx0, model->output, inpL);
|
|
|
|
|
assert_shape_2d(inpL, n_vocab, N*n_batch);
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
// inpL shape [n_vocab,N,n_batch,1]
|
|
|
|
|
inpL = ggml_reshape_3d(ctx0,
|
|
|
|
|
inpL,
|
|
|
|
|
n_vocab, N, n_batch);
|
|
|
|
|
assert_shape_3d(inpL, n_vocab, N, n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// run the computation
|
|
|
|
|
ggml_build_forward_expand(gf, inpL);
|
|
|
|
|
|
|
|
|
|
return inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * forward_batch_wo_cache(
|
|
|
|
|
struct my_llama_model * model,
|
|
|
|
|
struct ggml_context * ctx0,
|
|
|
|
|
struct ggml_cgraph * gf,
|
|
|
|
|
struct ggml_tensor * tokens_input,
|
|
|
|
|
const int n_tokens,
|
|
|
|
|
const int n_batch) {
|
|
|
|
|
|
|
|
|
|
const int n_past = 0;
|
|
|
|
|
const int N = n_tokens;
|
|
|
|
|
|
|
|
|
|
const auto & hparams = model->hparams;
|
|
|
|
|
//const int n_ctx = hparams.n_ctx;
|
|
|
|
|
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;
|
|
|
|
|
const int n_ff = get_n_ff(&hparams);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
|
|
|
|
|
memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1]
|
|
|
|
|
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cur;
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 0);
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = attention_norm*cur
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// self-attention
|
|
|
|
|
{
|
|
|
|
|
// compute Q and K and RoPE them
|
|
|
|
|
// wq shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// wk shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
// Kcur shape [n_embd/n_head, n_head, N, n_batch]
|
2023-06-26 23:37:13 +02:00
|
|
|
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
|
|
|
|
|
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 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
|
|
|
assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
|
|
|
|
|
// Vcur shape [N, n_batch, n_embd/n_head, n_head]
|
|
|
|
|
struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
|
|
|
|
|
// Qcur shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
// Q shape [n_embd/n_head, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * Q =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Qcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
|
|
|
|
|
// K shape [n_embd/n_head, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * K =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Kcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// K * Q
|
|
|
|
|
// KQ shape [N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
|
|
|
|
assert_shape_4d(KQ, N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
|
|
|
|
// KQ_scaled shape [N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_scaled =
|
|
|
|
|
ggml_scale_inplace(ctx0,
|
|
|
|
|
KQ,
|
|
|
|
|
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
|
|
|
|
|
assert_shape_4d(KQ_scaled, N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ_masked = mask_past(KQ_scaled)
|
|
|
|
|
// KQ_masked shape [N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
|
|
|
|
assert_shape_4d(KQ_masked, N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQ = soft_max(KQ_masked)
|
|
|
|
|
// KQ_soft_max shape [N, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
|
|
|
|
|
assert_shape_4d(KQ_soft_max, N, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// Vcur shape [N, n_batch, n_embd/n_head, n_head]
|
|
|
|
|
// V shape [N, n_embd/n_head, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * V =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Vcur,
|
|
|
|
|
0, 3, 1, 2);
|
|
|
|
|
assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQV shape [n_embd/n_head, N, n_head, n_batch]
|
|
|
|
|
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
|
|
|
|
|
assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
|
|
|
|
// KQV_merged shape [n_embd/n_head, n_head, N, n_batch]
|
|
|
|
|
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
// KQV_merged shape
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// projection (no bias)
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].wo,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lctx.use_buf(ctx0, 1);
|
|
|
|
|
|
|
|
|
|
// inpFF shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
|
|
|
|
|
assert_shape_2d(inpFF, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// feed-forward network
|
|
|
|
|
{
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = ffn_norm*cur
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// tmp shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w3,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(tmp, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w1,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// SILU activation
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_silu(ctx0, cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_ff,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul(ctx0, cur, tmp);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w2,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// cur shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
cur = ggml_add_inplace(ctx0, cur, inpFF);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// input for next layer
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
inpL = cur;
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
2023-07-24 17:57:12 +02:00
|
|
|
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// inpL = norm*inpL
|
|
|
|
|
// inpL shape [n_embd,N*n_batch,1,1]
|
|
|
|
|
inpL = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->norm, inpL),
|
|
|
|
|
inpL);
|
|
|
|
|
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
//embeddings = inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lm_head
|
|
|
|
|
// inpL shape [n_vocab,N*n_batch,1,1]
|
|
|
|
|
inpL = ggml_mul_mat(ctx0, model->output, inpL);
|
|
|
|
|
assert_shape_2d(inpL, n_vocab, N*n_batch);
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
// inpL shape [n_vocab,N,n_batch,1]
|
|
|
|
|
inpL = ggml_reshape_3d(ctx0,
|
|
|
|
|
inpL,
|
|
|
|
|
n_vocab, N, n_batch);
|
|
|
|
|
assert_shape_3d(inpL, n_vocab, N, n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// run the computation
|
|
|
|
|
ggml_build_forward_expand(gf, inpL);
|
|
|
|
|
|
|
|
|
|
return inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * forward_batch_wo_cache_flash_attn(
|
|
|
|
|
struct my_llama_model * model,
|
|
|
|
|
struct ggml_context * ctx0,
|
|
|
|
|
struct ggml_cgraph * gf,
|
|
|
|
|
struct ggml_tensor * tokens_input,
|
|
|
|
|
const int n_tokens,
|
|
|
|
|
const int n_batch) {
|
|
|
|
|
|
|
|
|
|
const int n_past = 0;
|
|
|
|
|
const int N = n_tokens;
|
|
|
|
|
|
|
|
|
|
const auto & hparams = model->hparams;
|
|
|
|
|
//const int n_ctx = hparams.n_ctx;
|
|
|
|
|
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;
|
|
|
|
|
const int n_ff = get_n_ff(&hparams);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
|
|
|
|
|
memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cur;
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = attention_norm*cur
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// self-attention
|
|
|
|
|
{
|
|
|
|
|
// compute Q and K and RoPE them
|
|
|
|
|
// wq shape [n_embd, n_embd, 1, 1]
|
|
|
|
|
// wk shape [n_embd, n_embd, 1, 1]
|
2023-06-26 23:37:13 +02:00
|
|
|
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
|
|
|
|
|
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 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
|
|
|
assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * Q =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Qcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * K =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Kcur,
|
|
|
|
|
0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * V =
|
|
|
|
|
ggml_permute(ctx0,
|
|
|
|
|
Vcur,
|
|
|
|
|
0, 3, 1, 2);
|
|
|
|
|
assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
bool masked = true;
|
|
|
|
|
struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, masked);
|
|
|
|
|
assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
|
|
|
|
assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// projection (no bias)
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].wo,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
|
|
|
|
|
assert_shape_2d(inpFF, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// feed-forward network
|
|
|
|
|
{
|
|
|
|
|
// norm
|
|
|
|
|
{
|
2023-07-24 17:57:12 +02:00
|
|
|
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_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
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// cur = ffn_norm*cur
|
|
|
|
|
cur = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w3,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(tmp, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w1,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// SILU activation
|
|
|
|
|
cur = ggml_silu(ctx0, cur);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
cur = ggml_mul(ctx0, cur, tmp);
|
|
|
|
|
assert_shape_2d(cur, n_ff, N*n_batch);
|
|
|
|
|
|
|
|
|
|
cur = ggml_mul_mat(ctx0,
|
|
|
|
|
model->layers[il].w2,
|
|
|
|
|
cur);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
cur = ggml_add_inplace(ctx0, cur, inpFF);
|
|
|
|
|
assert_shape_2d(cur, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// input for next layer
|
|
|
|
|
inpL = cur;
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// norm
|
|
|
|
|
{
|
|
|
|
|
|
2023-07-24 17:57:12 +02:00
|
|
|
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_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
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// inpL = norm*inpL
|
|
|
|
|
inpL = ggml_mul(ctx0,
|
|
|
|
|
ggml_repeat(ctx0, model->norm, inpL),
|
|
|
|
|
inpL);
|
|
|
|
|
|
|
|
|
|
assert_shape_2d(inpL, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lm_head
|
|
|
|
|
inpL = ggml_mul_mat(ctx0, model->output, inpL);
|
|
|
|
|
assert_shape_2d(inpL, n_vocab, N*n_batch);
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
inpL = ggml_reshape_3d(ctx0,
|
|
|
|
|
inpL,
|
|
|
|
|
n_vocab, N, n_batch);
|
|
|
|
|
assert_shape_3d(inpL, n_vocab, N, n_batch);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// run the computation
|
|
|
|
|
ggml_build_forward_expand(gf, inpL);
|
|
|
|
|
|
|
|
|
|
return inpL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// expand the graph nodes without creating leafs.
|
|
|
|
|
struct ggml_tensor * expand(struct ggml_cgraph * g, struct ggml_tensor * t) {
|
|
|
|
|
// check if already visited
|
|
|
|
|
for (int i = 0; i < g->n_nodes; i++) {
|
|
|
|
|
if (g->nodes[i] == t) {
|
|
|
|
|
return t;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (int i = 0; i < g->n_leafs; i++) {
|
|
|
|
|
if (g->leafs[i] == t) {
|
|
|
|
|
return t;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-11 18:31:10 +02:00
|
|
|
for (int i = 0; i < GGML_MAX_SRC; ++i) {
|
|
|
|
|
if (t->src[i]) {
|
|
|
|
|
expand(g, t->src[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
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
GGML_ASSERT(g->n_nodes < GGML_MAX_NODES);
|
|
|
|
|
|
|
|
|
|
if (strlen(t->name) == 0) {
|
|
|
|
|
snprintf(t->name, sizeof(t->name), "node_%d", g->n_nodes);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
g->nodes[g->n_nodes] = t;
|
|
|
|
|
g->grads[g->n_nodes] = t->grad;
|
|
|
|
|
g->n_nodes++;
|
|
|
|
|
return t;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void graph_set_leafs_grads(struct ggml_cgraph * g) {
|
|
|
|
|
// moves leaf nodes to g->leafs.
|
|
|
|
|
// i.e. g->n_nodes might change.
|
|
|
|
|
int n_nodes = 0;
|
|
|
|
|
for (int i = 0; i < g->n_nodes; ++i) {
|
|
|
|
|
struct ggml_tensor * node = g->nodes[i];
|
|
|
|
|
const bool is_leaf = node->op == GGML_OP_NONE && node->grad == NULL;
|
|
|
|
|
if (is_leaf) {
|
|
|
|
|
GGML_ASSERT(g->n_leafs < GGML_MAX_NODES);
|
|
|
|
|
|
|
|
|
|
if (strlen(node->name) == 0) {
|
|
|
|
|
snprintf(node->name, sizeof(node->name), "leaf_%d", g->n_leafs);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
g->leafs[g->n_leafs] = node;
|
|
|
|
|
g->n_leafs++;
|
|
|
|
|
} else {
|
|
|
|
|
GGML_ASSERT(n_nodes < GGML_MAX_NODES);
|
|
|
|
|
|
|
|
|
|
if (strlen(node->name) == 0) {
|
|
|
|
|
snprintf(node->name, sizeof(node->name), "node_%d", n_nodes);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
g->nodes[n_nodes] = node;
|
|
|
|
|
g->grads[n_nodes] = node->grad;
|
|
|
|
|
n_nodes++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
for (int i=n_nodes; i < g->n_nodes; ++i) {
|
|
|
|
|
g->nodes[n_nodes] = NULL;
|
|
|
|
|
g->grads[n_nodes] = NULL;
|
|
|
|
|
}
|
|
|
|
|
g->n_nodes = n_nodes;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
|
|
|
|
|
struct my_llama_model * model,
|
|
|
|
|
struct ggml_context * ctx0,
|
|
|
|
|
struct ggml_cgraph * gf,
|
|
|
|
|
struct ggml_cgraph * gb,
|
|
|
|
|
struct ggml_tensor * * logits,
|
|
|
|
|
struct ggml_tensor * tokens_input,
|
|
|
|
|
struct ggml_tensor * targets,
|
|
|
|
|
void * compute_buf_0,
|
|
|
|
|
void * compute_buf_1,
|
|
|
|
|
size_t size_buf_0,
|
|
|
|
|
size_t size_buf_1,
|
|
|
|
|
const int n_tokens,
|
|
|
|
|
const int n_batch) {
|
|
|
|
|
|
|
|
|
|
ggml_set_scratch(ctx0, { 0, 0, nullptr, });
|
|
|
|
|
|
|
|
|
|
const int n_past = 0;
|
|
|
|
|
const int N = n_tokens;
|
|
|
|
|
|
|
|
|
|
gf->n_nodes = 0;
|
|
|
|
|
gf->n_leafs = 0;
|
|
|
|
|
gf->perf_runs = 0;
|
|
|
|
|
gf->perf_cycles = 0;
|
|
|
|
|
gf->perf_time_us = 0;
|
|
|
|
|
|
|
|
|
|
const auto & hparams = model->hparams;
|
2023-07-21 13:51:34 +02:00
|
|
|
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;
|
|
|
|
|
const int n_ff = get_n_ff(&hparams);
|
|
|
|
|
const int rope_mode = 0;
|
|
|
|
|
|
|
|
|
|
int last_buf = -1;
|
|
|
|
|
size_t buf_offs[2] = { 0, 0 };
|
|
|
|
|
size_t buf_size[2] = { size_buf_0,
|
|
|
|
|
size_buf_1 };
|
|
|
|
|
void * buf_data[2] = { compute_buf_0,
|
|
|
|
|
compute_buf_1 };
|
|
|
|
|
auto use_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data] (int buf) {
|
|
|
|
|
size_t last_offs = 0;
|
|
|
|
|
last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, });
|
|
|
|
|
if (last_buf >= 0) {
|
|
|
|
|
buf_offs[last_buf] = last_offs;
|
|
|
|
|
}
|
|
|
|
|
if (buf >= 0) {
|
|
|
|
|
size_t offs = buf_offs[buf];
|
|
|
|
|
size_t size = buf_size[buf];
|
|
|
|
|
void * data = buf_data[buf];
|
|
|
|
|
ggml_set_scratch(ctx0, { offs, size, data, });
|
|
|
|
|
}
|
|
|
|
|
last_buf = buf;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
bool track_max_mem = false;
|
|
|
|
|
size_t buf_maxs[2] = { 0, 0 };
|
|
|
|
|
|
|
|
|
|
auto clr_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data, &buf_maxs, track_max_mem] (int buf) {
|
|
|
|
|
if (buf < 0) return;
|
|
|
|
|
if (track_max_mem) {
|
|
|
|
|
size_t last_offs = 0;
|
|
|
|
|
last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, });
|
|
|
|
|
if (last_buf >= 0) {
|
|
|
|
|
buf_offs[last_buf] = last_offs;
|
|
|
|
|
buf_maxs[last_buf] = std::max(buf_maxs[last_buf], buf_offs[last_buf]);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
buf_offs[buf] = 0;
|
|
|
|
|
if (track_max_mem && last_buf >= 0) {
|
|
|
|
|
size_t offs = buf_offs[last_buf];
|
|
|
|
|
size_t size = buf_size[last_buf];
|
|
|
|
|
void * data = buf_data[last_buf];
|
|
|
|
|
ggml_set_scratch(ctx0, { offs, size, data, });
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
auto view__q = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
|
|
|
|
|
int64_t ne0 = n_embd/n_head;
|
|
|
|
|
int64_t ne1 = N;
|
|
|
|
|
int64_t ne2 = n_head;
|
|
|
|
|
int64_t ne3 = n_batch;
|
|
|
|
|
size_t nb0 = ggml_element_size(t);
|
|
|
|
|
size_t nb1 = nb0*ne0;
|
|
|
|
|
size_t nb2 = nb1*ne1;
|
|
|
|
|
size_t nb3 = nb2*ne2;
|
|
|
|
|
size_t offset = 0;
|
|
|
|
|
return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
auto view__k = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
|
|
|
|
|
int64_t ne0 = n_embd/n_head;
|
|
|
|
|
int64_t ne1 = N;
|
|
|
|
|
int64_t ne2 = n_head;
|
|
|
|
|
int64_t ne3 = n_batch;
|
|
|
|
|
size_t nb0 = ggml_element_size(t);
|
|
|
|
|
size_t nb1 = nb0*ne0;
|
|
|
|
|
size_t nb2 = nb1*ne1;
|
|
|
|
|
size_t nb3 = nb2*ne2;
|
|
|
|
|
size_t offset = nb3*ne3;
|
|
|
|
|
return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
auto view__v = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
|
|
|
|
|
int64_t ne0 = N;
|
|
|
|
|
int64_t ne1 = n_embd/n_head;
|
|
|
|
|
int64_t ne2 = n_head;
|
|
|
|
|
int64_t ne3 = n_batch;
|
|
|
|
|
size_t nb0 = ggml_element_size(t);
|
|
|
|
|
size_t nb1 = nb0*ne0;
|
|
|
|
|
size_t nb2 = nb1*ne1;
|
|
|
|
|
size_t nb3 = nb2*ne2;
|
|
|
|
|
size_t offset = 2*nb3*ne3;
|
|
|
|
|
return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
auto add_or_set = [ctx0] (struct ggml_tensor * a, struct ggml_tensor * b) -> struct ggml_tensor * {
|
|
|
|
|
if (a == NULL) {
|
|
|
|
|
return b;
|
|
|
|
|
} else {
|
|
|
|
|
return ggml_add_inplace(ctx0, a, b);
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
|
|
|
|
|
model->tok_embeddings->grad = NULL;
|
|
|
|
|
model->norm->grad = NULL;
|
|
|
|
|
model->output->grad = NULL;
|
|
|
|
|
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
struct my_llama_layer & layer = model->layers[il];
|
|
|
|
|
layer.attention_norm->grad = NULL;
|
|
|
|
|
layer.wq->grad = NULL;
|
|
|
|
|
layer.wk->grad = NULL;
|
|
|
|
|
layer.wv->grad = NULL;
|
|
|
|
|
layer.wo->grad = NULL;
|
|
|
|
|
layer.ffn_norm->grad = NULL;
|
|
|
|
|
layer.w1->grad = NULL;
|
|
|
|
|
layer.w2->grad = NULL;
|
|
|
|
|
layer.w3->grad = NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
clr_buf(0);
|
|
|
|
|
clr_buf(1);
|
|
|
|
|
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * t00 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); assert_shape_1d(t00, N*n_batch);
|
|
|
|
|
memcpy(t00->data, tokens_input->data, ggml_element_size(t00)*N*n_batch);
|
|
|
|
|
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * t01 = expand(gf, ggml_get_rows(ctx0, model->tok_embeddings, t00)); assert_shape_2d(t01, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
// need to remember these for the backward pass
|
|
|
|
|
std::vector<struct ggml_tensor *> t02L; t02L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t03L; t03L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t04L; t04L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t05L; t05L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t06L; t06L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t07L; t07L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t08L; t08L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t09L; t09L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t10L; t10L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t11L; t11L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t12L; t12L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t13L; t13L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t14L; t14L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t15L; t15L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t16L; t16L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t17L; t17L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t18L; t18L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t19L; t19L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t20L; t20L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t21L; t21L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t22L; t22L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t23L; t23L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t24L; t24L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t25L; t25L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t26L; t26L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t27L; t27L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t28L; t28L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t29L; t29L.resize(n_layer, NULL);
|
|
|
|
|
std::vector<struct ggml_tensor *> t30L; t30L.resize(n_layer, NULL);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cur = t01;
|
|
|
|
|
|
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
|
|
|
clr_buf(0);
|
|
|
|
|
struct my_llama_layer & layer = model->layers[il];
|
|
|
|
|
// tensors with values necessary for backward pass are in persistent buf(-1)
|
|
|
|
|
// other tensors with buf(0) and buf(1) are only temporary needed, and their memory reused after layer is completed.
|
2023-07-24 17:57:12 +02:00
|
|
|
use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t02, 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
|
|
|
use_buf( 0); struct ggml_tensor * t03 = expand(gf, ggml_repeat (ctx0, layer.attention_norm, t02)); assert_shape_2d(t03, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t04 = expand(gf, ggml_mul (ctx0, t02, t03)); assert_shape_2d(t04, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t05 = expand(gf, ggml_mul_mat (ctx0, layer.wq, t04)); assert_shape_2d(t05, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t06 = expand(gf, ggml_reshape_4d (ctx0, t05, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch);
|
2023-06-26 23:37:13 +02:00
|
|
|
use_buf(-1); struct ggml_tensor * t07 = expand(gf, ggml_rope_inplace (ctx0, t06, n_past, n_rot, rope_mode, 0)); assert_shape_4d(t07, n_embd/n_head, n_head, 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
|
|
|
use_buf(-1); struct ggml_tensor * t08 = expand(gf, ggml_mul_mat (ctx0, layer.wk, t04)); assert_shape_2d(t08, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t09 = expand(gf, ggml_reshape_4d (ctx0, t08, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch);
|
2023-06-26 23:37:13 +02:00
|
|
|
use_buf(-1); struct ggml_tensor * t10 = expand(gf, ggml_rope_inplace (ctx0, t09, n_past, n_rot, rope_mode, 0)); assert_shape_4d(t10, n_embd/n_head, n_head, 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
|
|
|
use_buf(-1); struct ggml_tensor * t11 = expand(gf, ggml_mul_mat (ctx0, t04, layer.wv)); assert_shape_2d(t11, N*n_batch, n_embd);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t12 = expand(gf, ggml_reshape_4d (ctx0, t11, N, n_batch, n_embd/n_head, n_head)); assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t13 = expand(gf, ggml_permute (ctx0, t07, 0, 2, 1, 3)); assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t14 = expand(gf, ggml_permute (ctx0, t10, 0, 2, 1, 3)); assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t15 = expand(gf, ggml_permute (ctx0, t12, 0, 3, 1, 2)); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t16 = expand(gf, ggml_flash_attn (ctx0, t13, t14, t15, true)); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
use_buf( 0); struct ggml_tensor * t17 = expand(gf, ggml_permute (ctx0, t16, 0, 2, 1, 3)); assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t18 = expand(gf, ggml_cont (ctx0, t17)); assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t19 = expand(gf, ggml_reshape_2d (ctx0, t18, n_embd, N*n_batch)); assert_shape_2d(t19, n_embd, N*n_batch);
|
|
|
|
|
use_buf( 0); struct ggml_tensor * t20 = expand(gf, ggml_mul_mat (ctx0, layer.wo, t19)); assert_shape_2d(t20, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t21 = expand(gf, ggml_add (ctx0, t20, cur)); assert_shape_2d(t21, n_embd, N*n_batch);
|
2023-07-24 17:57:12 +02:00
|
|
|
use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm (ctx0, t21, rms_norm_eps)); assert_shape_2d(t22, 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
|
|
|
use_buf( 0); struct ggml_tensor * t23 = expand(gf, ggml_repeat (ctx0, layer.ffn_norm, t22)); assert_shape_2d(t23, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t24 = expand(gf, ggml_mul (ctx0, t23, t22)); assert_shape_2d(t24, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t25 = expand(gf, ggml_mul_mat (ctx0, layer.w3, t24)); assert_shape_2d(t25, n_ff, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t26 = expand(gf, ggml_mul_mat (ctx0, layer.w1, t24)); assert_shape_2d(t26, n_ff, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t27 = expand(gf, ggml_silu (ctx0, t26)); assert_shape_2d(t27, n_ff, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t28 = expand(gf, ggml_mul (ctx0, t27, t25)); assert_shape_2d(t28, n_ff, N*n_batch);
|
|
|
|
|
use_buf( 0); struct ggml_tensor * t29 = expand(gf, ggml_mul_mat (ctx0, layer.w2, t28)); assert_shape_2d(t29, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1); struct ggml_tensor * t30 = expand(gf, ggml_add (ctx0, t21, t29)); assert_shape_2d(t30, n_embd, N*n_batch);
|
|
|
|
|
t02L[il] = t02;
|
|
|
|
|
t03L[il] = t03;
|
|
|
|
|
t04L[il] = t04;
|
|
|
|
|
t05L[il] = t05;
|
|
|
|
|
t06L[il] = t06;
|
|
|
|
|
t07L[il] = t07;
|
|
|
|
|
t08L[il] = t08;
|
|
|
|
|
t09L[il] = t09;
|
|
|
|
|
t10L[il] = t10;
|
|
|
|
|
t11L[il] = t11;
|
|
|
|
|
t12L[il] = t12;
|
|
|
|
|
t13L[il] = t13;
|
|
|
|
|
t14L[il] = t14;
|
|
|
|
|
t15L[il] = t15;
|
|
|
|
|
t16L[il] = t16;
|
|
|
|
|
t17L[il] = t17;
|
|
|
|
|
t18L[il] = t18;
|
|
|
|
|
t19L[il] = t19;
|
|
|
|
|
t20L[il] = t20;
|
|
|
|
|
t21L[il] = t21;
|
|
|
|
|
t22L[il] = t22;
|
|
|
|
|
t23L[il] = t23;
|
|
|
|
|
t24L[il] = t24;
|
|
|
|
|
t25L[il] = t25;
|
|
|
|
|
t26L[il] = t26;
|
|
|
|
|
t27L[il] = t27;
|
|
|
|
|
t28L[il] = t28;
|
|
|
|
|
t29L[il] = t29;
|
|
|
|
|
t30L[il] = t30;
|
|
|
|
|
|
|
|
|
|
cur = t30;
|
|
|
|
|
}
|
|
|
|
|
clr_buf(0);
|
|
|
|
|
use_buf(0);
|
2023-07-24 17:57:12 +02:00
|
|
|
struct ggml_tensor * t31 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t31, 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
|
|
|
struct ggml_tensor * t32 = expand(gf, ggml_repeat (ctx0, model->norm, t31)); assert_shape_2d(t32, n_embd, N*n_batch);
|
|
|
|
|
struct ggml_tensor * t33 = expand(gf, ggml_mul (ctx0, t32, t31)); assert_shape_2d(t33, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
struct ggml_tensor * t34 = expand(gf, ggml_mul_mat (ctx0, model->output, t33)); assert_shape_2d(t34, n_vocab, N*n_batch);
|
|
|
|
|
struct ggml_tensor * t35 = expand(gf, ggml_reshape_3d(ctx0, t34, n_vocab, N, n_batch)); assert_shape_3d(t35, n_vocab, N, n_batch);
|
|
|
|
|
struct ggml_tensor * t36 = expand(gf, ggml_cross_entropy_loss(ctx0, t35, targets)); assert_shape_1d(t36, 1);
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
tok_embeddings | grad_tok_embeddings = ggml_get_rows_back(grad_t01, t00)
|
|
|
|
|
L0_att_norm | grad_L0_att_norm = ggml_repeat_back(grad_t03L0, L0_att_norm.shape)
|
|
|
|
|
L0_wq | grad_L0_wq = ggml_out_prod(t04L0, grad_t05L0)
|
|
|
|
|
L0_wk | grad_L0_wk = ggml_out_prod(t04L0, grad_t08L0)
|
|
|
|
|
L0_wv | grad_L0_wv = ggml_out_prod(t04L0, ggml_transpose(grad_t11L0))
|
|
|
|
|
L0_wo | grad_L0_wo = ggml_out_prod(t19L0, grad_t20L0)
|
|
|
|
|
L0_ffn_norm | grad_L0_ffn_norm = ggml_repeat_back(grad_t23L0, L0_ffn_norm.shape)
|
|
|
|
|
L0_w1 | grad_L0_w1 = ggml_out_prod(t24L0, grad_t26L0)
|
|
|
|
|
L0_w2 | grad_L0_w2 = ggml_out_prod(t28L0, grad_t29L0)
|
|
|
|
|
L0_w3 | grad_L0_w3 = ggml_out_prod(t24L0, grad_t25L0)
|
|
|
|
|
L1_att_norm | grad_L1_att_norm = ggml_repeat_back(grad_t03L1, L1_att_norm.shape)
|
|
|
|
|
L1_wq | grad_L1_wq = ggml_out_prod(t04L1, grad_t05L1)
|
|
|
|
|
L1_wk | grad_L1_wk = ggml_out_prod(t04L1, grad_t08L1)
|
|
|
|
|
L1_wv | grad_L1_wv = ggml_out_prod(t04L1, ggml_transpose(grad_t11L1))
|
|
|
|
|
L1_wo | grad_L1_wo = ggml_out_prod(t19L1, grad_t20L1)
|
|
|
|
|
L1_ffn_norm | grad_L1_ffn_norm = ggml_repeat_back(grad_t23L1, L1_ffn_norm.shape)
|
|
|
|
|
L1_w1 | grad_L1_w1 = ggml_out_prod(t24L1, grad_t26L1)
|
|
|
|
|
L1_w2 | grad_L1_w2 = ggml_out_prod(t28L1, grad_t29L1)
|
|
|
|
|
L1_w3 | grad_L1_w3 = ggml_out_prod(t24L1, grad_t25L1)
|
|
|
|
|
norm | grad_norm = ggml_repeat_back(grad_t32, norm.shape)
|
|
|
|
|
output | grad_output = ggml_out_prod(t33, grad_t34)
|
|
|
|
|
|
|
|
|
|
|
t01 = ggml_get_rows(tok_embeddings, t00) | grad_t01 = grad_t21L0 + ggml_rms_norm_back(t01, grad_t02L0)
|
|
|
|
|
for layer: |
|
|
|
|
|
t02L0*= ggml_rms_norm (t01) | grad_t02L0 = ggml_mul(grad_t04L0, t03L0)
|
|
|
|
|
t03L0 = ggml_repeat (L0_att_norm, t02L0_shape) | grad_t03L0 = ggml_mul(grad_t04L0, t02L0)
|
|
|
|
|
t04L0*= ggml_mul (t02L0, t03L0) | grad_t04L0 = ggml_out_prod(L0_wv, grad_t11L0) + ggml_out_prod(L0_wk, ggml_transpose(grad_t08L0)) + ggml_out_prod(L0_wq, ggml_transpose(grad_t05L0))
|
|
|
|
|
t05L0 = ggml_mul_mat (L0_wq, t04L0) | grad_t05L0 = ggml_reshape(grad_t06L0, t05L0_shape)
|
|
|
|
|
t06L0 = ggml_reshape_4d (t05L0, n_embd/n_head, n_head, N, n_batch) | grad_t06L0 = ggml_rope_back(grad_t07L0)
|
|
|
|
|
t07L0 = ggml_rope_inplace (t06L0) | grad_t07L0 = ggml_permute_back(grad_t13L0, 0, 2, 1, 3) = ggml_permute(grad_t13L0, 0, 2, 1, 3)
|
|
|
|
|
t08L0 = ggml_mul_mat (L0_wk, t04L0) | grad_t08L0 = ggml_reshape(grad_t09L0, t08L0_shape)
|
|
|
|
|
t09L0 = ggml_reshape_4d (t08L0, n_embd/n_head, n_head, N, n_batch) | grad_t09L0 = ggml_rope_back(grad_t10L0)
|
|
|
|
|
t10L0 = ggml_rope_inplace (t09L0) | grad_t10L0 = ggml_permute_back(grad_t14L0, 0, 2, 1, 3) = ggml_permute(grad_t14L0, 0, 2, 1, 3)
|
|
|
|
|
t11L0 = ggml_mul_mat (t04L0, L0_wv) | grad_t11L0 = ggml_reshape(grad_t12L0, t11L0_shape)
|
|
|
|
|
t12L0 = ggml_reshape_4d (t11L0, N, n_batch, n_embd/n_head, n_head) | grad_t12L0 = ggml_permute_back(grad_t15L0, 0, 3, 1, 2) = ggml_permute(grad_t15L0, 0, 2, 3, 1)
|
|
|
|
|
t13L0*= ggml_permute (t07L0, 0, 2, 1, 3) | grad_t13L0 = view__q(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
|
|
|
|
|
t14L0*= ggml_permute (t10L0, 0, 2, 1, 3) | grad_t14L0 = view__k(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
|
|
|
|
|
t15L0*= ggml_permute (t12L0, 0, 3, 1, 2) | grad_t15L0 = view__v(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
|
|
|
|
|
t16L0 = ggml_flash_attn (t13L0, t14L0, t15L0) | grad_t16L0 = ggml_permute_back(grad_t17L0, 0, 2, 1, 3) = ggml_permute(grad_t17L0, 0, 2, 1, 3)
|
|
|
|
|
t17L0 = ggml_permute (t16L0, 0, 2, 1, 3) | grad_t17L0 = grad_t18L0
|
|
|
|
|
t18L0 = ggml_cont (t17L0) | grad_t18L0 = ggml_reshape(grad_t19L0, t18L0_shape)
|
|
|
|
|
t19L0*= ggml_reshape_2d (t18L0, n_embd, N*n_batch) | grad_t19L0 = ggml_out_prod(L0_wo, ggml_transpose(grad_t20L0))
|
|
|
|
|
t20L0 = ggml_mul_mat (L0_wo, t19L0) | grad_t20L0 = grad_t21L0
|
|
|
|
|
t21L0*= ggml_add (t20L0, t01) | grad_t21L0 = grad_t30L0 + ggml_rms_norm_back(t21L0, grad_t22L0)
|
|
|
|
|
t22L0*= ggml_rms_norm (t21L0) | grad_t22L0 = ggml_mul(grad_t24L0, t23L0)
|
|
|
|
|
t23L0 = ggml_repeat (L0_ffn_norm, t22L0_shape) | grad_t23L0 = ggml_mul(grad_t24L0, t22L0)
|
|
|
|
|
t24L0*= ggml_mul (t23L0, t22L0) | grad_t24L0 = ggml_out_prod(L0_w1, ggml_transpose(grad_t26L0)) + ggml_out_prod(L0_w3, ggml_transpose(grad_t25L0))
|
|
|
|
|
t25L0*= ggml_mul_mat (L0_w3, t24L0) | grad_t25L0 = ggml_mul(grad_t28L0, t27L0)
|
|
|
|
|
t26L0*= ggml_mul_mat (L0_w1, t24L0) | grad_t26L0 = ggml_silu_back(t26L0, grad_t27L0)
|
|
|
|
|
t27L0*= ggml_silu (t26L0) | grad_t27L0 = ggml_mul(grad_t28L0, t25L0)
|
|
|
|
|
t28L0*= ggml_mul (t27L0, t25L0) | grad_t28L0 = ggml_out_prod(L0_w2, ggml_transpose(grad_t29L0))
|
|
|
|
|
t29L0 = ggml_mul_mat (L0_w2, t28L0) | grad_t29L0 = grad_t30L0
|
|
|
|
|
t30L0*= ggml_add (t21L0, t29L0) | grad_t30L0 = ggml_rms_norm_back(t30L0, grad_t02L1) + grad_t21L1
|
|
|
|
|
^
|
|
|
|
|
t02L1*= ggml_rms_norm (t30L0) | grad_t02L1 = ggml_mul(grad_t04L1, t03L1)
|
|
|
|
|
t03L1 = ggml_repeat (L1_att_norm, t02L1_shape) | grad_t03L1 = ggml_mul(grad_t04L1, t02L1)
|
|
|
|
|
t04L1*= ggml_mul (t02L1, t03L1) | grad_t04L1 = ggml_out_prod(L1_wv, grad_t11L1) + ggml_out_prod(L1_wk, ggml_transpose(grad_t08L1)) + ggml_out_prod(L1_wq, ggml_transpose(grad_t05L1))
|
|
|
|
|
t05L1 = ggml_mul_mat (L1_wq, t04L1) | grad_t05L1 = ggml_reshape(grad_t06L1, t05L1_shape)
|
|
|
|
|
t06L1 = ggml_reshape_4d (t05L1, n_embd/n_head, n_head, N, n_batch) | grad_t06L1 = ggml_rope_back(grad_t07L1)
|
|
|
|
|
t07L1 = ggml_rope_inplace (t06L1) | grad_t07L1 = ggml_permute_back(grad_t13L1, 0, 2, 1, 3) = ggml_permute(grad_t13L1, 0, 2, 1, 3)
|
|
|
|
|
t08L1 = ggml_mul_mat (L1_wk, t04L1) | grad_t08L1 = ggml_reshape(grad_t09L1, t08L1_shape)
|
|
|
|
|
t09L1 = ggml_reshape_4d (t08L1, n_embd/n_head, n_head, N, n_batch) | grad_t09L1 = ggml_rope_back(grad_t10L1)
|
|
|
|
|
t10L1 = ggml_rope_inplace (t09L1) | grad_t10L1 = ggml_permute_back(grad_t14L1, 0, 2, 1, 3) = ggml_permute(grad_t14L1, 0, 2, 1, 3)
|
|
|
|
|
t11L1 = ggml_mul_mat (t04L1, L1_wv) | grad_t11L1 = ggml_reshape(grad_t12L1, t11L1_shape)
|
|
|
|
|
t12L1 = ggml_reshape_4d (t11L1, N, n_batch, n_embd/n_head, n_head) | grad_t12L1 = ggml_permute_back(grad_t15L1, 0, 3, 1, 2) = ggml_permute(grad_t15L1, 0, 2, 3, 1)
|
|
|
|
|
t13L1*= ggml_permute (t07L1, 0, 2, 1, 3) | grad_t13L1 = view__q(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
|
|
|
|
|
t14L1*= ggml_permute (t10L1, 0, 2, 1, 3) | grad_t14L1 = view__k(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
|
|
|
|
|
t15L1*= ggml_permute (t12L1, 0, 3, 1, 2) | grad_t15L1 = view__v(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
|
|
|
|
|
t16L1 = ggml_flash_attn (t13L1, t14L1, t15L1) | grad_t16L1 = ggml_permute_back(grad_t17L1, 0, 2, 1, 3) = ggml_permute(grad_t17L1, 0, 2, 1, 3)
|
|
|
|
|
t17L1 = ggml_permute (t16L1, 0, 2, 1, 3) | grad_t17L1 = grad_t18L1
|
|
|
|
|
t18L1 = ggml_cont (t17L1) | grad_t18L1 = ggml_reshape(grad_t19L1, t18L1_shape)
|
|
|
|
|
t19L1*= ggml_reshape_2d (t18L1, n_embd, N*n_batch) | grad_t19L1 = ggml_out_prod(L1_wo, ggml_transpose(grad_t20L1))
|
|
|
|
|
t20L1 = ggml_mul_mat (L1_wo, t19L1) | grad_t20L1 = grad_t21L1
|
|
|
|
|
t21L1*= ggml_add (t20L1, t30L0) | grad_t21L1 = grad_t30L1 + ggml_rms_norm_back(t21L1, grad_t22L1)
|
|
|
|
|
t22L1*= ggml_rms_norm (t21L1) | grad_t22L1 = ggml_mul(grad_t24L1, t23L1)
|
|
|
|
|
t23L1 = ggml_repeat (L1_ffn_norm, t22L1_shape) | grad_t23L1 = ggml_mul(grad_t24L1, t22L1)
|
|
|
|
|
t24L1*= ggml_mul (t23L1, t22L1) | grad_t24L1 = ggml_out_prod(L1_w1, ggml_transpose(grad_t26L1)) + ggml_out_prod(L1_w3, ggml_transpose(grad_t25L1))
|
|
|
|
|
t25L1*= ggml_mul_mat (L1_w3, t24L1) | grad_t25L1 = ggml_mul(grad_t28L1, t27L1)
|
|
|
|
|
t26L1*= ggml_mul_mat (L1_w1, t24L1) | grad_t26L1 = ggml_silu_back(t26L1, grad_t27L1)
|
|
|
|
|
t27L1*= ggml_silu (t26L1) | grad_t27L1 = ggml_mul(grad_t28L1, t25L1)
|
|
|
|
|
t28L1*= ggml_mul (t27L1, t25L1) | grad_t28L1 = ggml_out_prod(L1_w2, ggml_transpose(grad_t29L1))
|
|
|
|
|
t29L1 = ggml_mul_mat (L1_w2, t28L1) | grad_t29L1 = grad_t30L1
|
|
|
|
|
t30L1*= ggml_add (t21L1, t29L1) | grad_t30L1 = ggml_rms_norm_back(t30L1, grad_t31)
|
|
|
|
|
^
|
|
|
|
|
t31 = ggml_rms_norm (t30L1) | grad_t31 = ggml_mul(grad_t33, t32)
|
|
|
|
|
t32 = ggml_repeat (norm, t31.shape) | grad_t32 = ggml_mul(grad_t33, t31)
|
|
|
|
|
t33 = ggml_mul (t32, t31) | grad_t33 = ggml_out_prod(output, ggml_transpose(grad_t34))
|
|
|
|
|
t34 = ggml_mul_mat (output, t33) | grad_t34 = ggml_reshape(grad_t35, t34.shape)
|
|
|
|
|
t35 = ggml_reshape_3d (t34, n_vocab, N, n_batch) | grad_t35 = ggml_cross_entropy_loss_back(t35, targets, grad_t36)
|
|
|
|
|
t36 = ggml_cross_entropy_loss(t35, targets) | grad_t36 = 1 (optimizer)
|
|
|
|
|
tensors marked with * need to be stored until grad computation
|
|
|
|
|
tensors during grad computation are all temporary
|
|
|
|
|
*/
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
*gb = *gf;
|
|
|
|
|
|
|
|
|
|
// t36->grad gets set to one by optimizer, so we need the tensor.
|
|
|
|
|
// initialize it with 1.0f to make sure.
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
t36->grad = expand(gb, ggml_new_f32(ctx0, 1.0f));
|
|
|
|
|
|
|
|
|
|
use_buf(0);
|
|
|
|
|
t35->grad = expand(gb, ggml_cross_entropy_loss_back(ctx0, t35, targets, t36->grad)); assert_shape_3d(t35->grad, n_vocab, N, n_batch);
|
|
|
|
|
t34->grad = expand(gb, ggml_reshape_2d (ctx0, t35->grad, n_vocab, N*n_batch)); assert_shape_2d(t34->grad, n_vocab, N*n_batch);
|
|
|
|
|
t33->grad = expand(gb, ggml_out_prod (ctx0, model->output, ggml_transpose(ctx0, t34->grad))); assert_shape_2d(t33->grad, n_embd, N*n_batch);
|
|
|
|
|
t32->grad = expand(gb, ggml_mul (ctx0, t33->grad, t31)); assert_shape_2d(t32->grad, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
|
|
|
|
|
model->norm->grad = expand(gb, add_or_set(model->norm->grad, ggml_repeat_back(ctx0, t32->grad, model->norm))); assert_shape_1d(model->norm->grad, n_embd);
|
|
|
|
|
model->output->grad = expand(gb, add_or_set(model->output->grad, ggml_out_prod(ctx0, t33, t34->grad))); assert_shape_2d(model->output->grad, n_embd, n_vocab);
|
|
|
|
|
|
|
|
|
|
clr_buf(1);
|
|
|
|
|
use_buf(1);
|
|
|
|
|
t31->grad = expand(gb, ggml_mul(ctx0, t33->grad, t32)); assert_shape_2d(t31->grad, n_embd, N*n_batch);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * back_layer_inp = t31;
|
|
|
|
|
struct ggml_tensor * grad_layer_inp = NULL;
|
|
|
|
|
|
|
|
|
|
for (int k = 0; k < n_layer; ++k) {
|
|
|
|
|
int il = n_layer-1-k;
|
|
|
|
|
struct my_llama_layer & layer = model->layers[il];
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * t02 = t02L[il];
|
|
|
|
|
struct ggml_tensor * t03 = t03L[il];
|
|
|
|
|
struct ggml_tensor * t04 = t04L[il];
|
|
|
|
|
struct ggml_tensor * t05 = t05L[il];
|
|
|
|
|
struct ggml_tensor * t06 = t06L[il];
|
|
|
|
|
struct ggml_tensor * t07 = t07L[il];
|
|
|
|
|
struct ggml_tensor * t08 = t08L[il];
|
|
|
|
|
struct ggml_tensor * t09 = t09L[il];
|
|
|
|
|
struct ggml_tensor * t10 = t10L[il];
|
|
|
|
|
struct ggml_tensor * t11 = t11L[il];
|
|
|
|
|
struct ggml_tensor * t12 = t12L[il];
|
|
|
|
|
struct ggml_tensor * t13 = t13L[il];
|
|
|
|
|
struct ggml_tensor * t14 = t14L[il];
|
|
|
|
|
struct ggml_tensor * t15 = t15L[il];
|
|
|
|
|
struct ggml_tensor * t16 = t16L[il];
|
|
|
|
|
struct ggml_tensor * t17 = t17L[il];
|
|
|
|
|
struct ggml_tensor * t18 = t18L[il];
|
|
|
|
|
struct ggml_tensor * t19 = t19L[il];
|
|
|
|
|
struct ggml_tensor * t20 = t20L[il];
|
|
|
|
|
struct ggml_tensor * t21 = t21L[il];
|
|
|
|
|
struct ggml_tensor * t22 = t22L[il];
|
|
|
|
|
struct ggml_tensor * t23 = t23L[il];
|
|
|
|
|
struct ggml_tensor * t24 = t24L[il];
|
|
|
|
|
struct ggml_tensor * t25 = t25L[il];
|
|
|
|
|
struct ggml_tensor * t26 = t26L[il];
|
|
|
|
|
struct ggml_tensor * t27 = t27L[il];
|
|
|
|
|
struct ggml_tensor * t28 = t28L[il];
|
|
|
|
|
struct ggml_tensor * t29 = t29L[il];
|
|
|
|
|
struct ggml_tensor * t30 = t30L[il];
|
|
|
|
|
|
|
|
|
|
clr_buf(0);
|
|
|
|
|
use_buf(0);
|
|
|
|
|
t30->grad = expand(gb, ggml_rms_norm_back(ctx0, t30, back_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch);
|
|
|
|
|
if (grad_layer_inp) {
|
|
|
|
|
t30->grad = expand(gb, ggml_add(ctx0, t30->grad, grad_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch);
|
|
|
|
|
}
|
|
|
|
|
clr_buf(1);
|
|
|
|
|
t29->grad = t30->grad; assert_shape_2d(t29->grad, n_embd, N*n_batch);
|
|
|
|
|
t28->grad = expand(gb, ggml_out_prod(ctx0, layer.w2, ggml_transpose(ctx0, t29->grad))); assert_shape_2d(t28->grad, n_ff, N*n_batch);
|
|
|
|
|
t27->grad = expand(gb, ggml_mul(ctx0, t28->grad, t25)); assert_shape_2d(t27->grad, n_ff, N*n_batch);
|
|
|
|
|
t26->grad = expand(gb, ggml_silu_back(ctx0, t26, t27->grad)); assert_shape_2d(t26->grad, n_ff, N*n_batch);
|
|
|
|
|
t25->grad = expand(gb, ggml_mul(ctx0, t28->grad, t27)); assert_shape_2d(t25->grad, n_ff, N*n_batch);
|
|
|
|
|
t24->grad = expand(gb, ggml_add_inplace(ctx0,
|
|
|
|
|
ggml_out_prod(ctx0, layer.w1, ggml_transpose(ctx0, t26->grad)),
|
|
|
|
|
ggml_out_prod(ctx0, layer.w3, ggml_transpose(ctx0, t25->grad)))); assert_shape_2d(t24->grad, n_embd, N*n_batch);
|
|
|
|
|
t23->grad = expand(gb, ggml_mul(ctx0, t24->grad, t22)); assert_shape_2d(t23->grad, n_embd, N*n_batch);
|
|
|
|
|
t22->grad = expand(gb, ggml_mul(ctx0, t24->grad, ggml_repeat(ctx0, layer.ffn_norm, t24->grad))); assert_shape_2d(t22->grad, n_embd, N*n_batch);
|
|
|
|
|
use_buf(1);
|
|
|
|
|
t21->grad = expand(gb, ggml_add(ctx0, t30->grad, ggml_rms_norm_back(ctx0, t21, t22->grad))); assert_shape_2d(t21->grad, n_embd, N*n_batch);
|
|
|
|
|
grad_layer_inp = t21;
|
|
|
|
|
use_buf(0);
|
|
|
|
|
t20->grad = t21->grad; assert_shape_2d(t20->grad, n_embd, N*n_batch);
|
|
|
|
|
t19->grad = expand(gb, ggml_out_prod(ctx0, layer.wo, ggml_transpose(ctx0, t20->grad))); assert_shape_2d(t19->grad, n_embd, N*n_batch);
|
|
|
|
|
t18->grad = expand(gb, ggml_reshape_4d(ctx0, t19->grad, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t18->grad, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
t17->grad = t18->grad; assert_shape_4d(t17->grad, n_embd/n_head, n_head, N, n_batch);
|
|
|
|
|
t16->grad = expand(gb, ggml_permute(ctx0, t17->grad, 0, 2, 1, 3)); assert_shape_4d(t16->grad, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
struct ggml_tensor * flash_attn = expand(gb, ggml_flash_attn_back(ctx0, t13, t14, t15, t16->grad, true)); assert_shape_4d(flash_attn, n_embd/n_head, N*3, n_head, n_batch);
|
|
|
|
|
t15->grad = expand(gb, view__v(flash_attn)); assert_shape_4d(t15->grad, N, n_embd/n_head, n_head, n_batch);
|
|
|
|
|
t14->grad = expand(gb, view__k(flash_attn)); assert_shape_4d(t14->grad, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
t13->grad = expand(gb, view__q(flash_attn)); assert_shape_4d(t13->grad, n_embd/n_head, N, n_head, n_batch);
|
|
|
|
|
t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1)); assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
|
|
|
|
|
t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd)); assert_shape_2d(t11->grad, N*n_batch, n_embd);
|
|
|
|
|
t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3)); assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
|
2023-08-22 13:22:08 +02:00
|
|
|
t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false)); assert_shape_4d(t09->grad, n_embd/n_head, n_head, 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
|
|
|
t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch)); assert_shape_2d(t08->grad, n_embd, N*n_batch);
|
|
|
|
|
t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3)); assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
|
2023-08-22 13:22:08 +02:00
|
|
|
t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false)); assert_shape_4d(t06->grad, n_embd/n_head, n_head, 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
|
|
|
t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch)); assert_shape_2d(t05->grad, n_embd, N*n_batch);
|
|
|
|
|
t04->grad = expand(gb, ggml_add_inplace(ctx0,
|
|
|
|
|
ggml_add_inplace(ctx0,
|
|
|
|
|
ggml_out_prod(ctx0, layer.wv, t11->grad),
|
|
|
|
|
ggml_out_prod(ctx0, layer.wk, ggml_transpose(ctx0, t08->grad))),
|
|
|
|
|
ggml_out_prod(ctx0, layer.wq, ggml_transpose(ctx0, t05->grad)))); assert_shape_2d(t04->grad, n_embd, N*n_batch);
|
|
|
|
|
t03->grad = expand(gb, ggml_mul(ctx0, t04->grad, t02)); assert_shape_2d(t04->grad, n_embd, N*n_batch);
|
|
|
|
|
use_buf(1);
|
|
|
|
|
t02->grad = expand(gb, ggml_mul(ctx0, t04->grad, ggml_repeat(ctx0, layer.attention_norm, t02))); assert_shape_2d(t02->grad, n_embd, N*n_batch);
|
|
|
|
|
back_layer_inp = t02;
|
|
|
|
|
// use_buf(0);
|
|
|
|
|
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
layer.attention_norm->grad = expand(gb, add_or_set(layer.attention_norm->grad, ggml_repeat_back(ctx0, t03->grad, layer.attention_norm))); assert_shape_1d(layer.attention_norm->grad, n_embd);
|
|
|
|
|
layer.wq->grad = expand(gb, add_or_set(layer.wq->grad, ggml_out_prod(ctx0, t04, t05->grad))); assert_shape_2d(layer.wq->grad, n_embd, n_embd);
|
|
|
|
|
layer.wk->grad = expand(gb, add_or_set(layer.wk->grad, ggml_out_prod(ctx0, t04, t08->grad))); assert_shape_2d(layer.wk->grad, n_embd, n_embd);
|
|
|
|
|
layer.wv->grad = expand(gb, add_or_set(layer.wv->grad, ggml_out_prod(ctx0, t04, ggml_transpose(ctx0, t11->grad)))); assert_shape_2d(layer.wv->grad, n_embd, n_embd);
|
|
|
|
|
layer.wo->grad = expand(gb, add_or_set(layer.wo->grad, ggml_out_prod(ctx0, t19, t20->grad))); assert_shape_2d(layer.wo->grad, n_embd, n_embd);
|
|
|
|
|
layer.ffn_norm->grad = expand(gb, add_or_set(layer.ffn_norm->grad, ggml_repeat_back(ctx0, t23->grad, layer.ffn_norm))); assert_shape_1d(layer.ffn_norm->grad, n_embd);
|
|
|
|
|
layer.w1->grad = expand(gb, add_or_set(layer.w1->grad, ggml_out_prod(ctx0, t24, t26->grad))); assert_shape_2d(layer.w1->grad, n_embd, n_ff);
|
|
|
|
|
layer.w2->grad = expand(gb, add_or_set(layer.w2->grad, ggml_out_prod(ctx0, t28, t29->grad))); assert_shape_2d(layer.w2->grad, n_ff, n_embd);
|
|
|
|
|
layer.w3->grad = expand(gb, add_or_set(layer.w3->grad, ggml_out_prod(ctx0, t24, t25->grad))); assert_shape_2d(layer.w3->grad, n_embd, n_ff);
|
|
|
|
|
// use_buf(0);
|
|
|
|
|
}
|
|
|
|
|
clr_buf(0);
|
|
|
|
|
use_buf(0);
|
|
|
|
|
t01->grad = expand(gb, ggml_add_inplace(ctx0, grad_layer_inp->grad, ggml_rms_norm_back(ctx0, t01, back_layer_inp->grad))); assert_shape_2d(t01->grad, n_embd, N*n_batch);
|
|
|
|
|
use_buf(-1);
|
|
|
|
|
model->tok_embeddings->grad = expand(gb, ggml_get_rows_back(ctx0, t01->grad, t00, model->tok_embeddings)); assert_shape_2d(model->tok_embeddings->grad, n_embd, n_vocab);
|
|
|
|
|
// clr_buf(1);
|
|
|
|
|
// clr_buf(0);
|
|
|
|
|
|
|
|
|
|
*logits = t35;
|
|
|
|
|
|
|
|
|
|
if (track_max_mem) {
|
|
|
|
|
printf("%s: max size compute buf0: %zu\n", __func__, buf_maxs[0]);
|
|
|
|
|
printf("%s: max size compute buf1: %zu\n", __func__, buf_maxs[1]);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// now that all grads are created, set the graph leafs and grads
|
|
|
|
|
graph_set_leafs_grads(gf);
|
|
|
|
|
graph_set_leafs_grads(gb);
|
|
|
|
|
|
|
|
|
|
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");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
void print_token(struct llama_context * ctx, llama_token token) {
|
2023-08-27 13:19:19 +02:00
|
|
|
printf("%s", llama_token_to_piece(ctx, token).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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) {
|
|
|
|
|
for (int i=0; i<tokens->ne[0]; ++i) {
|
|
|
|
|
int token = ggml_get_i32_1d(tokens, i);
|
|
|
|
|
print_token(ctx, token);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void print_tokens_batch(struct llama_context* ctx, struct ggml_tensor * tokens) {
|
|
|
|
|
for (int i1=0; i1<tokens->ne[1]; ++i1) {
|
|
|
|
|
//int num_newline = 0;
|
|
|
|
|
for (int i0=0; i0<tokens->ne[0]; ++i0) {
|
|
|
|
|
int token = get_i32_2d(tokens, i0, i1);
|
|
|
|
|
print_token(ctx, token);
|
|
|
|
|
// bool isnl = (token == llama_token_nl());
|
|
|
|
|
// if (isnl) {
|
|
|
|
|
// ++num_newline;
|
|
|
|
|
// }
|
|
|
|
|
// if (isnl) {
|
|
|
|
|
// if (num_newline < 2) {
|
|
|
|
|
// print_token(ctx, token);
|
|
|
|
|
// } else {
|
|
|
|
|
// printf("\\n");
|
|
|
|
|
// }
|
|
|
|
|
// } else {
|
|
|
|
|
// print_token(ctx, token);
|
|
|
|
|
// }
|
|
|
|
|
}
|
|
|
|
|
printf("\n--\n");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
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);
|
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);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
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);
|
|
|
|
|
for (int k=0; k<n_batch; ++k) {
|
|
|
|
|
// printf("%s: batch %d\n", __func__, k);
|
|
|
|
|
size_t sample = train_samples[(example_id*n_batch + k) % n_train_samples];
|
|
|
|
|
GGML_ASSERT(sample+n_tokens-1 < n_train_data);
|
|
|
|
|
|
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);
|
|
|
|
|
// print_token(lctx, token);
|
|
|
|
|
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);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
// printf("\n=\n");
|
|
|
|
|
// for (int i=0; i<n_tokens; ++i) {
|
|
|
|
|
// int token = get_i32_2d(tokens_input, i, k);
|
|
|
|
|
// print_token(lctx, token);
|
|
|
|
|
// }
|
|
|
|
|
// printf("\n-\n");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs, int n_shift) {
|
|
|
|
|
int n_tokens = tokens_input->ne[0];
|
|
|
|
|
int n_vocab = target_logits->ne[0];
|
|
|
|
|
for (int i=0; i<n_tokens-n_shift; ++i) {
|
|
|
|
|
ggml_set_i32_1d(tokens_input, i, ggml_get_i32_1d(tokens_input, i + n_shift));
|
|
|
|
|
for (int k=0; k<n_vocab; ++k) {
|
|
|
|
|
ggml_set_f32_1d(target_logits, i*n_vocab + k, ggml_get_f32_1d(target_logits, (i + n_shift)*n_vocab + k));
|
|
|
|
|
ggml_set_f32_1d(target_probs, i*n_vocab + k, ggml_get_f32_1d(target_probs, (i + n_shift)*n_vocab + k));
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * target) {
|
|
|
|
|
return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, target, a)));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * probs) {
|
|
|
|
|
return ggml_cross_entropy_loss(ctx, a, probs);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#ifdef __GNUC__
|
|
|
|
|
#ifdef __MINGW32__
|
|
|
|
|
__attribute__((format(gnu_printf, 1, 2)))
|
|
|
|
|
#else
|
|
|
|
|
__attribute__((format(printf, 1, 2)))
|
|
|
|
|
#endif
|
|
|
|
|
#endif
|
|
|
|
|
static std::string format(const char * fmt, ...) {
|
|
|
|
|
va_list ap, ap2;
|
|
|
|
|
va_start(ap, fmt);
|
|
|
|
|
va_copy(ap2, ap);
|
|
|
|
|
int size = vsnprintf(NULL, 0, fmt, ap);
|
|
|
|
|
GGML_ASSERT(size >= 0 && size < INT_MAX);
|
|
|
|
|
std::vector<char> buf(size + 1);
|
|
|
|
|
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
|
|
|
|
|
GGML_ASSERT(size2 == size);
|
|
|
|
|
va_end(ap2);
|
|
|
|
|
va_end(ap);
|
|
|
|
|
return std::string(buf.data(), size);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct llama_file {
|
|
|
|
|
// use FILE * so we don't have to re-open the file to mmap
|
|
|
|
|
FILE * fp;
|
|
|
|
|
size_t size;
|
|
|
|
|
|
|
|
|
|
llama_file(const char * fname, const char * mode) {
|
|
|
|
|
fp = std::fopen(fname, mode);
|
|
|
|
|
if (fp == NULL) {
|
|
|
|
|
size = 0;
|
|
|
|
|
} else {
|
|
|
|
|
seek(0, SEEK_END);
|
|
|
|
|
size = tell();
|
|
|
|
|
seek(0, SEEK_SET);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
size_t tell() const {
|
|
|
|
|
#ifdef _WIN32
|
|
|
|
|
__int64 ret = _ftelli64(fp);
|
|
|
|
|
#else
|
|
|
|
|
long ret = std::ftell(fp);
|
|
|
|
|
#endif
|
|
|
|
|
GGML_ASSERT(ret != -1); // this really shouldn't fail
|
|
|
|
|
return (size_t) ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void seek(size_t offset, int whence) {
|
|
|
|
|
#ifdef _WIN32
|
|
|
|
|
int ret = _fseeki64(fp, (__int64) offset, whence);
|
|
|
|
|
#else
|
|
|
|
|
int ret = std::fseek(fp, (long) offset, whence);
|
|
|
|
|
#endif
|
|
|
|
|
GGML_ASSERT(ret == 0); // same
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void read_raw(void * ptr, size_t size) {
|
|
|
|
|
if (size == 0) {
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
errno = 0;
|
|
|
|
|
std::size_t ret = std::fread(ptr, size, 1, fp);
|
|
|
|
|
if (ferror(fp)) {
|
|
|
|
|
throw std::runtime_error(format("read error: %s", strerror(errno)));
|
|
|
|
|
}
|
|
|
|
|
if (ret != 1) {
|
|
|
|
|
throw std::runtime_error(std::string("unexpectedly reached end of file"));
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
std::uint32_t read_u32() {
|
|
|
|
|
std::uint32_t ret;
|
|
|
|
|
read_raw(&ret, sizeof(ret));
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
std::string read_string(std::uint32_t len) {
|
|
|
|
|
std::vector<char> chars(len);
|
|
|
|
|
read_raw(chars.data(), len);
|
|
|
|
|
return std::string(chars.data(), len);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void write_raw(const void * ptr, size_t size) {
|
|
|
|
|
if (size == 0) {
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
errno = 0;
|
|
|
|
|
size_t ret = std::fwrite(ptr, size, 1, fp);
|
|
|
|
|
if (ret != 1) {
|
|
|
|
|
throw std::runtime_error(format("write error: %s", strerror(errno)));
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void write_u32(std::uint32_t val) {
|
|
|
|
|
write_raw(&val, sizeof(val));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
~llama_file() {
|
|
|
|
|
if (fp) {
|
|
|
|
|
std::fclose(fp);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
int tokenize_file(struct llama_context * lctx, const char * filename, std::vector<llama_token>& out) {
|
|
|
|
|
struct llama_file f(filename, "rb");
|
|
|
|
|
|
|
|
|
|
std::vector<char> buf;
|
|
|
|
|
buf.resize(f.size+1);
|
|
|
|
|
|
|
|
|
|
f.read_raw(buf.data(), f.size);
|
|
|
|
|
buf[f.size] = '\0';
|
|
|
|
|
|
2023-08-21 22:07:43 +02:00
|
|
|
int n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false);
|
|
|
|
|
if (n_tokens < 0) {
|
|
|
|
|
out.resize(-n_tokens);
|
|
|
|
|
llama_tokenize(lctx, buf.data(), 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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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) {
|
2023-08-27 13:19:19 +02:00
|
|
|
std::string s = llama_token_to_piece(lctx, out[i]);
|
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;
|
|
|
|
|
}
|
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 {
|
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);
|
|
|
|
|
});
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct my_llama_sampler_params {
|
|
|
|
|
float temp = 0.0f; // <= 0.0 disabled
|
|
|
|
|
int top_k = 20; // <= 0 to use vocab size
|
|
|
|
|
float top_p = 0.95f; // 1.0 = disabled
|
|
|
|
|
float tfs_z = 1.00f; // 1.0 = disabled
|
|
|
|
|
float typical_p = 1.00f; // 1.0 = disabled
|
|
|
|
|
int repeat_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
|
|
|
|
|
float repeat_penalty = 1.0f; // 1.0 = disabled
|
|
|
|
|
float alpha_presence = 0.0f; // 0.0 = disabled
|
|
|
|
|
float alpha_frequency = 0.0f; // 0.0 = disabled
|
|
|
|
|
int mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
|
|
|
|
|
float mirostat_tau = 5.00f; // target entropy
|
|
|
|
|
float mirostat_eta = 0.10f; // learning rate
|
|
|
|
|
bool penalize_nl = true; // consider newlines as a repeatable token
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
struct my_llama_sampler {
|
|
|
|
|
struct llama_context * ctx = NULL;
|
|
|
|
|
my_llama_sampler_params params;
|
|
|
|
|
|
|
|
|
|
int n_vocab = 0;
|
|
|
|
|
int n_ctx = 0;
|
|
|
|
|
|
|
|
|
|
float mirostat_mu;
|
|
|
|
|
|
|
|
|
|
std::vector<llama_token_data> candidates;
|
|
|
|
|
llama_token_data_array candidates_p;
|
|
|
|
|
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
void init_sampler(struct my_llama_sampler * sampler, struct llama_context * ctx) {
|
|
|
|
|
sampler->ctx = ctx;
|
|
|
|
|
sampler->n_vocab = llama_n_vocab(sampler->ctx);
|
|
|
|
|
sampler->n_ctx = llama_n_ctx(sampler->ctx);
|
|
|
|
|
sampler->mirostat_mu = 2.0f * sampler->params.mirostat_tau;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
llama_token sample(struct my_llama_sampler * sampler, float * logits, const llama_token * last_tokens, int n_last_tokens) {
|
|
|
|
|
GGML_ASSERT(sampler->ctx != NULL);
|
|
|
|
|
|
|
|
|
|
struct llama_context * ctx = sampler->ctx;
|
|
|
|
|
|
|
|
|
|
sampler->candidates.resize(sampler->n_vocab);
|
|
|
|
|
for (llama_token token_id = 0; token_id < sampler->n_vocab; ++token_id) {
|
|
|
|
|
sampler->candidates[token_id].id = token_id;
|
|
|
|
|
sampler->candidates[token_id].logit = logits[token_id];
|
|
|
|
|
sampler->candidates[token_id].p = 0.0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
llama_token_data_array * candidates_p = & sampler->candidates_p;
|
|
|
|
|
|
|
|
|
|
candidates_p->data = sampler->candidates.data();
|
|
|
|
|
candidates_p->size = sampler->candidates.size();
|
|
|
|
|
candidates_p->sorted = false;
|
|
|
|
|
|
|
|
|
|
const auto params = sampler->params;
|
|
|
|
|
|
|
|
|
|
// Apply penalties
|
2023-08-21 22:07:43 +02:00
|
|
|
const float nl_logit = logits[llama_token_nl(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_last = std::min(std::min(n_last_tokens, params.repeat_last_n), sampler->n_ctx);
|
|
|
|
|
|
|
|
|
|
llama_sample_repetition_penalty(
|
|
|
|
|
ctx,
|
|
|
|
|
candidates_p,
|
|
|
|
|
last_tokens + n_last_tokens - n_last,
|
|
|
|
|
n_last,
|
|
|
|
|
params.repeat_penalty);
|
|
|
|
|
llama_sample_frequency_and_presence_penalties(
|
|
|
|
|
ctx,
|
|
|
|
|
candidates_p,
|
|
|
|
|
last_tokens + n_last_tokens - n_last,
|
|
|
|
|
n_last,
|
|
|
|
|
params.alpha_frequency,
|
|
|
|
|
params.alpha_presence);
|
|
|
|
|
|
|
|
|
|
if (!params.penalize_nl) {
|
2023-08-21 22:07:43 +02:00
|
|
|
logits[llama_token_nl(ctx)] = nl_logit;
|
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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
llama_token token = 0;
|
|
|
|
|
if (params.temp <= 0) {
|
|
|
|
|
// Greedy sampling
|
|
|
|
|
token = llama_sample_token_greedy(ctx, candidates_p);
|
|
|
|
|
} else {
|
|
|
|
|
if (params.mirostat == 1) {
|
|
|
|
|
int mirostat_m = 100;
|
|
|
|
|
llama_sample_temperature(ctx, candidates_p, params.temp);
|
|
|
|
|
token = llama_sample_token_mirostat(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, mirostat_m, &sampler->mirostat_mu);
|
|
|
|
|
} else if (params.mirostat == 2) {
|
|
|
|
|
llama_sample_temperature(ctx, candidates_p, params.temp);
|
|
|
|
|
token = llama_sample_token_mirostat_v2(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, &sampler->mirostat_mu);
|
|
|
|
|
} else {
|
|
|
|
|
// Temperature sampling
|
|
|
|
|
llama_sample_top_k (ctx, candidates_p, params.top_k, 1);
|
|
|
|
|
llama_sample_tail_free (ctx, candidates_p, params.tfs_z, 1);
|
|
|
|
|
llama_sample_typical (ctx, candidates_p, params.typical_p, 1);
|
|
|
|
|
|
|
|
|
|
llama_sample_top_p (ctx, candidates_p, params.top_p, 1);
|
|
|
|
|
llama_sample_temperature (ctx, candidates_p, params.temp);
|
|
|
|
|
token = llama_sample_token(ctx, candidates_p);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return token;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void set_logits_masked(struct ggml_tensor * logits, std::vector<bool>& mask, float value) {
|
|
|
|
|
GGML_ASSERT(logits->ne[0] == (int64_t) mask.size());
|
|
|
|
|
for (int i2 = 0; i2 < logits->ne[2]; ++i2) {
|
|
|
|
|
for (int i1 = 0; i1 < logits->ne[1]; ++i1) {
|
|
|
|
|
for (int i0 = 0; i0 < logits->ne[0]; ++i0) {
|
|
|
|
|
if (!mask[i0]) continue;
|
|
|
|
|
float * ptr = (float *) ((char *) logits->data + i2*logits->nb[2] + i1*logits->nb[1] + i0*logits->nb[0]);
|
|
|
|
|
*ptr = value;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
|
|
|
|
|
if (tensor == NULL) {
|
|
|
|
|
file->write_u32(0);
|
|
|
|
|
file->write_u32(0);
|
|
|
|
|
file->write_u32(GGML_TYPE_F32);
|
2023-06-26 21:45:32 +02:00
|
|
|
file->seek((0-file->tell()) & 31, SEEK_CUR);
|
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;
|
|
|
|
|
}
|
|
|
|
|
const char * name = ggml_get_name(tensor);
|
|
|
|
|
uint32_t name_len = strlen(name);
|
|
|
|
|
uint32_t nd = tensor->n_dims;
|
|
|
|
|
uint32_t ne[4] = { (uint32_t)tensor->ne[0],
|
|
|
|
|
(uint32_t)tensor->ne[1],
|
|
|
|
|
(uint32_t)tensor->ne[2],
|
|
|
|
|
(uint32_t)tensor->ne[3] };
|
|
|
|
|
file->write_u32(nd);
|
|
|
|
|
file->write_u32(name_len);
|
|
|
|
|
file->write_u32(tensor->type);
|
|
|
|
|
file->write_raw(ne, sizeof(ne[0]) * nd);
|
|
|
|
|
file->write_raw(name, name_len);
|
2023-06-26 21:45:32 +02:00
|
|
|
file->seek((0-file->tell()) & 31, SEEK_CUR);
|
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
|
|
|
file->write_raw(tensor->data, ggml_nbytes(tensor));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void read_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
|
|
|
|
|
int32_t nd = file->read_u32();
|
|
|
|
|
GGML_ASSERT(nd == tensor->n_dims);
|
|
|
|
|
|
|
|
|
|
uint32_t name_len = file->read_u32();
|
|
|
|
|
enum ggml_type type = (enum ggml_type) file->read_u32();
|
|
|
|
|
GGML_ASSERT(type == tensor->type);
|
|
|
|
|
|
|
|
|
|
uint32_t ne[4];
|
|
|
|
|
file->read_raw(ne, sizeof(ne[0]) * nd);
|
|
|
|
|
for (int i=0; i<nd; ++i) {
|
|
|
|
|
GGML_ASSERT(ne[i] == tensor->ne[i]);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
std::string name = file->read_string(name_len);
|
|
|
|
|
GGML_ASSERT(strncmp(ggml_get_name(tensor), name.c_str(), sizeof(tensor->name)-1) == 0);
|
|
|
|
|
|
2023-06-26 21:45:32 +02:00
|
|
|
file->seek((0-file->tell()) & 31, SEEK_CUR);
|
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
|
|
|
file->read_raw(tensor->data, ggml_nbytes(tensor));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void write_opt_context(struct llama_file * file, struct ggml_opt_context * opt) {
|
|
|
|
|
const uint32_t version = 0;
|
|
|
|
|
GGML_ASSERT(opt->nx >= 0);
|
|
|
|
|
GGML_ASSERT(opt->iter >= 0);
|
|
|
|
|
file->write_u32(version);
|
|
|
|
|
file->write_raw(&opt->params, sizeof(opt->params));
|
|
|
|
|
file->write_raw(&opt->nx, sizeof(opt->nx));
|
|
|
|
|
file->write_raw(&opt->iter, sizeof(opt->iter));
|
|
|
|
|
file->write_u32((uint32_t) opt->just_initialized);
|
|
|
|
|
switch (opt->params.type) {
|
|
|
|
|
case GGML_OPT_ADAM:
|
|
|
|
|
{
|
|
|
|
|
GGML_ASSERT(opt->adam.x != NULL);
|
|
|
|
|
write_tensor(file, opt->adam.x);
|
|
|
|
|
write_tensor(file, opt->adam.g1);
|
|
|
|
|
write_tensor(file, opt->adam.g2);
|
|
|
|
|
write_tensor(file, opt->adam.m);
|
|
|
|
|
write_tensor(file, opt->adam.v);
|
|
|
|
|
write_tensor(file, opt->adam.mh);
|
|
|
|
|
write_tensor(file, opt->adam.vh);
|
|
|
|
|
write_tensor(file, opt->adam.pf);
|
|
|
|
|
file->write_raw(&opt->adam.fx_best, sizeof(opt->adam.fx_best));
|
|
|
|
|
file->write_raw(&opt->adam.fx_prev, sizeof(opt->adam.fx_prev));
|
|
|
|
|
file->write_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement));
|
|
|
|
|
} break;
|
|
|
|
|
case GGML_OPT_LBFGS:
|
|
|
|
|
{
|
|
|
|
|
GGML_ASSERT(opt->adam.x != NULL);
|
|
|
|
|
write_tensor(file, opt->lbfgs.x);
|
|
|
|
|
write_tensor(file, opt->lbfgs.xp);
|
|
|
|
|
write_tensor(file, opt->lbfgs.g);
|
|
|
|
|
write_tensor(file, opt->lbfgs.gp);
|
|
|
|
|
write_tensor(file, opt->lbfgs.d);
|
|
|
|
|
write_tensor(file, opt->lbfgs.pf);
|
|
|
|
|
write_tensor(file, opt->lbfgs.lmal);
|
|
|
|
|
write_tensor(file, opt->lbfgs.lmys);
|
|
|
|
|
write_tensor(file, opt->lbfgs.lms);
|
|
|
|
|
write_tensor(file, opt->lbfgs.lmy);
|
|
|
|
|
file->write_raw(&opt->lbfgs.fx_best, sizeof(opt->lbfgs.fx_best));
|
|
|
|
|
file->write_raw(&opt->lbfgs.step, sizeof(opt->lbfgs.step));
|
|
|
|
|
file->write_raw(&opt->lbfgs.j, sizeof(opt->lbfgs.j));
|
|
|
|
|
file->write_raw(&opt->lbfgs.k, sizeof(opt->lbfgs.k));
|
|
|
|
|
file->write_raw(&opt->lbfgs.end, sizeof(opt->lbfgs.end));
|
|
|
|
|
file->write_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement));
|
|
|
|
|
} break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void read_opt_context(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) {
|
|
|
|
|
uint32_t version = file->read_u32();
|
|
|
|
|
GGML_ASSERT(version == 0);
|
|
|
|
|
|
|
|
|
|
file->read_raw(&opt->params, sizeof(opt->params));
|
|
|
|
|
file->read_raw(&opt->nx, sizeof(opt->nx));
|
|
|
|
|
ggml_opt_init(ctx, opt, opt->params, opt->nx);
|
|
|
|
|
|
|
|
|
|
file->read_raw(&opt->iter, sizeof(opt->iter));
|
|
|
|
|
opt->just_initialized = (bool) file->read_u32();
|
|
|
|
|
|
|
|
|
|
switch (opt->params.type) {
|
|
|
|
|
case GGML_OPT_ADAM:
|
|
|
|
|
{
|
|
|
|
|
read_tensor(file, opt->adam.x);
|
|
|
|
|
read_tensor(file, opt->adam.g1);
|
|
|
|
|
read_tensor(file, opt->adam.g2);
|
|
|
|
|
read_tensor(file, opt->adam.m);
|
|
|
|
|
read_tensor(file, opt->adam.v);
|
|
|
|
|
read_tensor(file, opt->adam.mh);
|
|
|
|
|
read_tensor(file, opt->adam.vh);
|
|
|
|
|
if (opt->adam.pf) { read_tensor(file, opt->adam.pf); }
|
|
|
|
|
file->read_raw(&opt->adam.fx_best, sizeof(opt->adam.fx_best));
|
|
|
|
|
file->read_raw(&opt->adam.fx_prev, sizeof(opt->adam.fx_prev));
|
|
|
|
|
file->read_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement));
|
|
|
|
|
} break;
|
|
|
|
|
case GGML_OPT_LBFGS:
|
|
|
|
|
{
|
|
|
|
|
GGML_ASSERT(opt->adam.x != NULL);
|
|
|
|
|
read_tensor(file, opt->lbfgs.x);
|
|
|
|
|
read_tensor(file, opt->lbfgs.xp);
|
|
|
|
|
read_tensor(file, opt->lbfgs.g);
|
|
|
|
|
read_tensor(file, opt->lbfgs.gp);
|
|
|
|
|
read_tensor(file, opt->lbfgs.d);
|
|
|
|
|
if (opt->lbfgs.pf) { read_tensor(file, opt->lbfgs.pf); }
|
|
|
|
|
read_tensor(file, opt->lbfgs.lmal);
|
|
|
|
|
read_tensor(file, opt->lbfgs.lmys);
|
|
|
|
|
read_tensor(file, opt->lbfgs.lms);
|
|
|
|
|
read_tensor(file, opt->lbfgs.lmy);
|
|
|
|
|
file->read_raw(&opt->lbfgs.fx_best, sizeof(opt->lbfgs.fx_best));
|
|
|
|
|
file->read_raw(&opt->lbfgs.step, sizeof(opt->lbfgs.step));
|
|
|
|
|
file->read_raw(&opt->lbfgs.j, sizeof(opt->lbfgs.j));
|
|
|
|
|
file->read_raw(&opt->lbfgs.k, sizeof(opt->lbfgs.k));
|
|
|
|
|
file->read_raw(&opt->lbfgs.end, sizeof(opt->lbfgs.end));
|
|
|
|
|
file->read_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement));
|
|
|
|
|
} break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void save_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename) {
|
|
|
|
|
struct llama_file file(filename, "wb");
|
|
|
|
|
if (file.fp == NULL) {
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const uint32_t magic = 'ggcp';
|
|
|
|
|
const uint32_t version = 0;
|
|
|
|
|
|
|
|
|
|
file.write_u32(magic);
|
|
|
|
|
file.write_u32(version);
|
|
|
|
|
file.write_u32(model->train_its);
|
|
|
|
|
file.write_u32(model->train_samples);
|
|
|
|
|
file.write_u32(model->train_tokens);
|
|
|
|
|
file.write_u32(model->hparams.n_vocab);
|
|
|
|
|
file.write_u32(model->hparams.n_embd);
|
|
|
|
|
file.write_u32(model->hparams.n_mult);
|
|
|
|
|
file.write_u32(model->hparams.n_head);
|
|
|
|
|
file.write_u32(model->hparams.n_layer);
|
|
|
|
|
file.write_u32(model->hparams.n_rot);
|
|
|
|
|
|
|
|
|
|
write_tensor(&file, model->tok_embeddings);
|
|
|
|
|
write_tensor(&file, model->norm);
|
|
|
|
|
write_tensor(&file, model->output);
|
|
|
|
|
|
|
|
|
|
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
|
|
|
|
|
auto & layer = model->layers[i];
|
|
|
|
|
|
|
|
|
|
write_tensor(&file, layer.attention_norm);
|
|
|
|
|
write_tensor(&file, layer.wq);
|
|
|
|
|
write_tensor(&file, layer.wk);
|
|
|
|
|
write_tensor(&file, layer.wv);
|
|
|
|
|
write_tensor(&file, layer.wo);
|
|
|
|
|
write_tensor(&file, layer.ffn_norm);
|
|
|
|
|
write_tensor(&file, layer.w1);
|
|
|
|
|
write_tensor(&file, layer.w2);
|
|
|
|
|
write_tensor(&file, layer.w3);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
write_opt_context(&file, opt);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
bool load_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename, bool init) {
|
|
|
|
|
struct llama_file file(filename, "rb");
|
|
|
|
|
|
|
|
|
|
uint32_t magic;
|
|
|
|
|
uint32_t version;
|
|
|
|
|
|
|
|
|
|
uint32_t train_its = 0;
|
|
|
|
|
uint32_t train_samples = 0;
|
|
|
|
|
uint32_t train_tokens = 0;
|
|
|
|
|
|
|
|
|
|
if (file.fp) {
|
|
|
|
|
printf("%s: Loading model from '%s'.\n", __func__, filename);
|
|
|
|
|
magic = file.read_u32();
|
|
|
|
|
GGML_ASSERT(magic == 'ggcp');
|
|
|
|
|
version = file.read_u32();
|
|
|
|
|
GGML_ASSERT(version == 0);
|
|
|
|
|
train_its = file.read_u32();
|
|
|
|
|
train_samples = file.read_u32();
|
|
|
|
|
train_tokens = file.read_u32();
|
|
|
|
|
model->hparams.n_vocab = file.read_u32();
|
|
|
|
|
model->hparams.n_embd = file.read_u32();
|
|
|
|
|
model->hparams.n_mult = file.read_u32();
|
|
|
|
|
model->hparams.n_head = file.read_u32();
|
|
|
|
|
model->hparams.n_layer = file.read_u32();
|
|
|
|
|
model->hparams.n_rot = file.read_u32();
|
|
|
|
|
print_params(&model->hparams);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (init) {
|
|
|
|
|
init_model(model);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (file.fp) {
|
|
|
|
|
model->train_its = train_its;
|
|
|
|
|
model->train_samples = train_samples;
|
|
|
|
|
model->train_tokens = train_tokens;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
printf("%s: Training iterations: %u.\n", __func__, model->train_its);
|
|
|
|
|
printf("%s: Training samples: %u.\n", __func__, model->train_samples);
|
|
|
|
|
printf("%s: Training tokens: %u.\n", __func__, model->train_tokens);
|
|
|
|
|
|
|
|
|
|
if (file.fp) {
|
|
|
|
|
read_tensor(&file, model->tok_embeddings);
|
|
|
|
|
read_tensor(&file, model->norm);
|
|
|
|
|
read_tensor(&file, model->output);
|
|
|
|
|
|
|
|
|
|
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
|
|
|
|
|
auto & layer = model->layers[i];
|
|
|
|
|
|
|
|
|
|
read_tensor(&file, layer.attention_norm);
|
|
|
|
|
read_tensor(&file, layer.wq);
|
|
|
|
|
read_tensor(&file, layer.wk);
|
|
|
|
|
read_tensor(&file, layer.wv);
|
|
|
|
|
read_tensor(&file, layer.wo);
|
|
|
|
|
read_tensor(&file, layer.ffn_norm);
|
|
|
|
|
read_tensor(&file, layer.w1);
|
|
|
|
|
read_tensor(&file, layer.w2);
|
|
|
|
|
read_tensor(&file, layer.w3);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
read_opt_context(&file, model->ctx, opt);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return (file.fp != NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, const char * filename) {
|
|
|
|
|
struct llama_file file(filename, "wb");
|
|
|
|
|
if (file.fp == NULL) {
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-21 22:07:43 +02:00
|
|
|
#pragma message("TODO: implement file saving using gguf")
|
|
|
|
|
(void) vocab;
|
|
|
|
|
(void) model;
|
|
|
|
|
// // write_magic
|
|
|
|
|
// file.write_u32(LLAMA_FILE_MAGIC); // magic
|
|
|
|
|
// file.write_u32(LLAMA_FILE_VERSION); // version
|
|
|
|
|
// // write_hparams
|
|
|
|
|
// file.write_u32(model->hparams.n_vocab);
|
|
|
|
|
// file.write_u32(model->hparams.n_embd);
|
|
|
|
|
// file.write_u32(model->hparams.n_mult);
|
|
|
|
|
// file.write_u32(model->hparams.n_head);
|
|
|
|
|
// file.write_u32(model->hparams.n_layer);
|
|
|
|
|
// file.write_u32(model->hparams.n_rot);
|
|
|
|
|
// file.write_u32(LLAMA_FTYPE_ALL_F32);
|
|
|
|
|
// // write_vocab
|
|
|
|
|
// uint32_t n_vocab = model->hparams.n_vocab;
|
|
|
|
|
// for (uint32_t i = 0; i < n_vocab; i++) {
|
|
|
|
|
// const auto & token_data = vocab->id_to_token.at(i);
|
|
|
|
|
// file.write_u32((uint32_t) token_data.tok.size());
|
|
|
|
|
// file.write_raw(token_data.tok.data(), token_data.tok.size());
|
|
|
|
|
// file.write_raw(&token_data.score, sizeof(token_data.score));
|
|
|
|
|
// }
|
|
|
|
|
// // write tensors
|
|
|
|
|
// write_tensor(&file, model->tok_embeddings);
|
|
|
|
|
// write_tensor(&file, model->norm);
|
|
|
|
|
// write_tensor(&file, model->output);
|
|
|
|
|
// for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
|
|
|
|
|
// auto & layer = model->layers[i];
|
|
|
|
|
//
|
|
|
|
|
// write_tensor(&file, layer.attention_norm);
|
|
|
|
|
// write_tensor(&file, layer.wq);
|
|
|
|
|
// write_tensor(&file, layer.wk);
|
|
|
|
|
// write_tensor(&file, layer.wv);
|
|
|
|
|
// write_tensor(&file, layer.wo);
|
|
|
|
|
// write_tensor(&file, layer.ffn_norm);
|
|
|
|
|
// write_tensor(&file, layer.w1);
|
|
|
|
|
// write_tensor(&file, layer.w2);
|
|
|
|
|
// write_tensor(&file, 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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
float cosine_decay(const int decay_steps, const float alpha, int step) {
|
|
|
|
|
if (step > decay_steps) {
|
|
|
|
|
step = decay_steps;
|
|
|
|
|
}
|
|
|
|
|
const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps));
|
|
|
|
|
const float decay = (1 - alpha)*cosine_decay + alpha;
|
|
|
|
|
return decay;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
float cosine_decay_restart(int decay_steps, const float alpha, int step, float restart_step_mult) {
|
|
|
|
|
while (step > decay_steps) {
|
|
|
|
|
step -= decay_steps;
|
|
|
|
|
decay_steps = (int) restart_step_mult * decay_steps;
|
|
|
|
|
}
|
|
|
|
|
return cosine_decay(decay_steps, alpha, step);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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_mult;
|
|
|
|
|
int n_head;
|
|
|
|
|
int n_layer;
|
|
|
|
|
int n_rotmax;
|
|
|
|
|
|
|
|
|
|
int n_threads;
|
|
|
|
|
int n_batch;
|
|
|
|
|
int n_examples;
|
|
|
|
|
int n_predict;
|
|
|
|
|
|
|
|
|
|
int print_info_interval;
|
|
|
|
|
int print_details_interval;
|
|
|
|
|
|
|
|
|
|
bool samples_start_after_nl;
|
|
|
|
|
bool use_adam;
|
|
|
|
|
bool use_flash;
|
|
|
|
|
bool use_scratch;
|
|
|
|
|
|
|
|
|
|
// only adam
|
|
|
|
|
int warmup;
|
|
|
|
|
int cos_decay_steps;
|
|
|
|
|
float cos_decay_restart;
|
|
|
|
|
float cos_decay_alpha;
|
|
|
|
|
|
|
|
|
|
int lbfgs_n_iter;
|
|
|
|
|
int adam_n_iter;
|
|
|
|
|
float adam_alpha;
|
|
|
|
|
float adam_decay;
|
|
|
|
|
|
|
|
|
|
int mem_model_gb;
|
|
|
|
|
int mem_compute_gb;
|
|
|
|
|
int mem_compute0_gb;
|
|
|
|
|
int mem_compute1_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_mult = 256;
|
|
|
|
|
params.n_head = 8;
|
|
|
|
|
params.n_layer = 16;
|
|
|
|
|
params.n_rotmax = 64;
|
|
|
|
|
|
|
|
|
|
params.n_threads = 6;
|
|
|
|
|
params.n_batch = 8;
|
|
|
|
|
params.n_examples = 8;
|
|
|
|
|
params.n_predict = 1024;
|
|
|
|
|
|
|
|
|
|
params.print_info_interval = 1;
|
|
|
|
|
params.print_details_interval = 2;
|
|
|
|
|
|
|
|
|
|
params.samples_start_after_nl = false;
|
|
|
|
|
params.use_adam = true;
|
|
|
|
|
params.use_flash = true;
|
|
|
|
|
params.use_scratch = true;
|
|
|
|
|
|
|
|
|
|
// only adam
|
|
|
|
|
params.warmup = 100;
|
|
|
|
|
params.cos_decay_steps = 1000;
|
|
|
|
|
params.cos_decay_restart = 1.1f;
|
|
|
|
|
params.cos_decay_alpha = 0.0f;
|
|
|
|
|
|
|
|
|
|
params.lbfgs_n_iter = 16;
|
|
|
|
|
params.adam_n_iter = 16;
|
2023-06-16 20:23:53 +02:00
|
|
|
params.adam_alpha = 1e-3f;
|
|
|
|
|
params.adam_decay = 1e-3f;
|
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_model_gb = 2;
|
|
|
|
|
params.mem_compute_gb = 24;
|
|
|
|
|
params.mem_compute0_gb = 8;
|
|
|
|
|
params.mem_compute1_gb = 2;
|
|
|
|
|
|
|
|
|
|
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);
|
2023-06-29 15:15:15 +02:00
|
|
|
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);
|
|
|
|
|
fprintf(stderr, " --mult N Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult);
|
|
|
|
|
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);
|
|
|
|
|
fprintf(stderr, " --rotmax N Maximal number Rope dimensions for new models (default %d)\n", params->n_rotmax);
|
|
|
|
|
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, " --predict N Number of tokens to generate after training (default %d)\n", params->n_predict);
|
|
|
|
|
fprintf(stderr, " --print-info-interval N Print infos during training each N examples (default %d)\n", params->print_info_interval);
|
|
|
|
|
fprintf(stderr, " --print-details-interval N Print details during training each N examples (default %d)\n", params->print_details_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");
|
|
|
|
|
fprintf(stderr, " --no-flash Don't use flash attention.\n");
|
|
|
|
|
fprintf(stderr, " --use-flash Use flash attention (default)\n");
|
|
|
|
|
fprintf(stderr, " --no-scratch Don't use scratch buffers\n");
|
|
|
|
|
fprintf(stderr, " --use-scratch Use scratch buffers (default)\n");
|
|
|
|
|
fprintf(stderr, " --warmup N Number of warmup steps (default %d)\n", params->warmup);
|
|
|
|
|
fprintf(stderr, " --cos-decay-steps N Number of cosine decay steps (default %d)\n", params->cos_decay_steps);
|
|
|
|
|
fprintf(stderr, " --cos-decay-restart N Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart);
|
|
|
|
|
fprintf(stderr, " --cos-decay-alpha N Cosine decay alpha (default %f)\n", params->cos_decay_alpha);
|
|
|
|
|
fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter);
|
|
|
|
|
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);
|
|
|
|
|
fprintf(stderr, " --adam-decay N AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
|
|
|
|
|
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);
|
|
|
|
|
fprintf(stderr, " --mem-compute0 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute0_gb);
|
|
|
|
|
fprintf(stderr, " --mem-compute1 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute1_gb);
|
|
|
|
|
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]);
|
|
|
|
|
} else if (arg == "--mult") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_mult = std::stoi(argv[i]);
|
|
|
|
|
} 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]);
|
|
|
|
|
} else if (arg == "--rotmax") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_rotmax = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "-t" || arg == "--threads") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_threads = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "-b" || arg == "--batch") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_batch = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "-n" || arg == "--examples") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_examples = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "--predict") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->n_predict = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "--print-info-interval") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->print_info_interval = std::stoi(argv[i]);
|
|
|
|
|
} else if (arg == "--print-details-interval") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->print_details_interval = std::stoi(argv[i]);
|
|
|
|
|
} 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;
|
|
|
|
|
} else if (arg == "--no-scratch") {
|
|
|
|
|
params->use_scratch = false;
|
|
|
|
|
} else if (arg == "--use-scratch") {
|
|
|
|
|
params->use_scratch = true;
|
|
|
|
|
} 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]);
|
|
|
|
|
} else if (arg == "--cos-decay-alpha") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->cos_decay_alpha = std::stof(argv[i]);
|
|
|
|
|
} else if (arg == "--lbfgs-iter") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->lbfgs_n_iter = std::stoi(argv[i]);
|
|
|
|
|
} 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]);
|
|
|
|
|
} else if (arg == "--adam-decay") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->adam_decay = std::stof(argv[i]);
|
|
|
|
|
} 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 == "--mem-compute1") {
|
|
|
|
|
if (++i >= argc) {
|
|
|
|
|
invalid_param = true;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
params->mem_compute1_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;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
int main(int argc, char ** argv) {
|
|
|
|
|
struct train_params params = get_default_train_params();
|
|
|
|
|
|
|
|
|
|
if (!train_params_parse(argc, argv, ¶ms)) {
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
2023-06-29 15:15:15 +02:00
|
|
|
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);
|
|
|
|
|
}
|
2023-06-29 15:15:15 +02:00
|
|
|
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;
|
|
|
|
|
|
2023-06-24 10:47:58 +02:00
|
|
|
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
|
|
|
|
|
|
|
|
struct llama_vocab vocab;
|
|
|
|
|
{
|
2023-08-21 22:07:43 +02:00
|
|
|
const int n_vocab = llama_n_vocab(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
|
|
|
vocab.id_to_token.resize(n_vocab);
|
|
|
|
|
for (int i=0; i<n_vocab; ++i) {
|
2023-08-21 22:07:43 +02:00
|
|
|
vocab.id_to_token[i].text = llama_token_get_text(lctx, i);
|
|
|
|
|
vocab.id_to_token[i].score = llama_token_get_score(lctx, i);
|
|
|
|
|
vocab.id_to_token[i].type = llama_token_get_type(lctx, i);
|
|
|
|
|
vocab.token_to_id.emplace(vocab.id_to_token[i].text, 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
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
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_mult = params.n_mult;
|
|
|
|
|
model.hparams.n_head = params.n_head;
|
|
|
|
|
model.hparams.n_layer = params.n_layer;
|
|
|
|
|
model.hparams.n_rot = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
|
|
|
|
|
|
|
|
|
|
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 my_llama_kv_cache kv_self;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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);
|
|
|
|
|
kv_self.ctx = model.ctx;
|
|
|
|
|
|
|
|
|
|
my_llama_sampler sampler;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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);
|
|
|
|
|
opt_params_adam.print_forward_graph = false;
|
|
|
|
|
opt_params_adam.print_backward_graph = false;
|
|
|
|
|
opt_params_adam.n_threads = params.n_threads;
|
|
|
|
|
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_lbfgs.print_forward_graph = false;
|
|
|
|
|
opt_params_lbfgs.print_backward_graph = false;
|
|
|
|
|
opt_params_lbfgs.n_threads = params.n_threads;
|
|
|
|
|
opt_params_lbfgs.lbfgs.n_iter = params.lbfgs_n_iter;
|
|
|
|
|
|
|
|
|
|
opt->ctx = model.ctx;
|
|
|
|
|
opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs;
|
|
|
|
|
|
|
|
|
|
printf("%s: init model\n", __func__);
|
|
|
|
|
bool existed = load_checkpoint(&model, opt, params.fn_checkpoint_in, true);
|
|
|
|
|
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);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
init_kv_cache(&kv_self, &model, 1);
|
|
|
|
|
// init_kv_cache(&kv_self, &model, n_batch);
|
|
|
|
|
init_sampler(&sampler, lctx);
|
|
|
|
|
|
|
|
|
|
printf("used_mem model+cache: %zu bytes\n", ggml_used_mem(model.ctx));
|
|
|
|
|
// ggml_print_tensor_objects(model.ctx);
|
|
|
|
|
|
2023-07-07 18:24:01 +02:00
|
|
|
// 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);
|
|
|
|
|
size_t size_buf_1 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute1_gb);
|
|
|
|
|
uint8_t * compute_buf_0 = new uint8_t[size_buf_0];
|
|
|
|
|
uint8_t * compute_buf_1 = new uint8_t[size_buf_1];
|
|
|
|
|
|
|
|
|
|
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) {
|
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());
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-07 18:24:01 +02:00
|
|
|
std::vector<uint8_t> work_buffer;
|
|
|
|
|
|
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: begin training\n", __func__);
|
|
|
|
|
|
|
|
|
|
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 = {
|
|
|
|
|
/*.mem_size =*/ compute_size,
|
|
|
|
|
/*.mem_buffer =*/ compute_addr,
|
|
|
|
|
/*.no_alloc =*/ false,
|
|
|
|
|
};
|
|
|
|
|
struct ggml_context * ctx0 = ggml_init(cparams);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * after_opt_best_samples = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch);
|
|
|
|
|
//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);
|
|
|
|
|
|
|
|
|
|
int n_past = 0;
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * gfbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0));
|
|
|
|
|
struct ggml_tensor * gbbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0));
|
|
|
|
|
|
|
|
|
|
memset(gfbuf->data, 0, ggml_nbytes(gfbuf));
|
|
|
|
|
memset(gbbuf->data, 0, ggml_nbytes(gbbuf));
|
|
|
|
|
|
|
|
|
|
struct ggml_cgraph * gf = (struct ggml_cgraph *) gfbuf->data;
|
|
|
|
|
struct ggml_cgraph * gb = (struct ggml_cgraph *) gbbuf->data;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get_example_targets_batch(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), ex, tokens_input, target_logits, target_probs);
|
|
|
|
|
|
|
|
|
|
GGML_ASSERT(n_past == 0);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * loss = NULL;
|
|
|
|
|
struct ggml_tensor * logits = NULL;
|
|
|
|
|
|
|
|
|
|
if (params.use_scratch) {
|
|
|
|
|
loss = forward_batch_wo_cache_flash_attn_train(
|
|
|
|
|
&model, ctx0,
|
|
|
|
|
gf, gb,
|
|
|
|
|
&logits, tokens_input, target_probs,
|
|
|
|
|
compute_buf_0, compute_buf_1,
|
|
|
|
|
size_buf_0, size_buf_1,
|
|
|
|
|
n_tokens, n_batch);
|
|
|
|
|
} else if (params.use_flash) {
|
|
|
|
|
logits = forward_batch_wo_cache_flash_attn(&model, ctx0, gf, tokens_input, n_tokens, n_batch);
|
|
|
|
|
loss = cross_entropy_loss(ctx0, logits, target_probs);
|
|
|
|
|
ggml_build_forward_expand(gf, loss);
|
|
|
|
|
*gb = ggml_build_backward(ctx0, gf, true);
|
|
|
|
|
} else {
|
|
|
|
|
logits = forward_batch_wo_cache(&model, ctx0, gf, tokens_input, n_tokens, n_batch);
|
|
|
|
|
loss = cross_entropy_loss(ctx0, logits, target_probs);
|
|
|
|
|
ggml_build_forward_expand(gf, loss);
|
|
|
|
|
*gb = ggml_build_backward(ctx0, gf, true);
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-07 18:24:01 +02:00
|
|
|
ggml_graph_compute_helper(work_buffer, gf, params.n_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
|
|
|
|
|
|
|
|
size_t used_mem_before_opt = ggml_used_mem(ctx0);
|
|
|
|
|
|
|
|
|
|
float error_before_opt = ggml_get_f32_1d(loss, 0);
|
|
|
|
|
|
|
|
|
|
opt->params.adam.sched = (opt->iter < params.warmup)
|
|
|
|
|
? (float) opt->iter / (float) params.warmup
|
|
|
|
|
: cosine_decay_restart(
|
|
|
|
|
params.cos_decay_steps,
|
|
|
|
|
params.cos_decay_alpha,
|
|
|
|
|
opt->iter - params.warmup,
|
|
|
|
|
params.cos_decay_restart);
|
|
|
|
|
|
|
|
|
|
printf("%s: opt->params.adam.sched %.5f\n", __func__, opt->params.adam.sched);
|
|
|
|
|
|
|
|
|
|
ggml_opt_resume_g(ctx0, opt, loss, gf, gb);
|
|
|
|
|
|
|
|
|
|
size_t used_mem_after_opt = ggml_used_mem(ctx0);
|
|
|
|
|
|
|
|
|
|
model.train_its = opt->iter;
|
|
|
|
|
model.train_samples += n_batch;
|
|
|
|
|
model.train_tokens += n_batch * n_tokens;
|
|
|
|
|
|
2023-07-07 18:24:01 +02:00
|
|
|
ggml_graph_compute_helper(work_buffer, gf, params.n_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
|
|
|
|
|
|
|
|
float error_after_opt = ggml_get_f32_1d(loss, 0);
|
|
|
|
|
|
|
|
|
|
if (params.print_info_interval > 0 && ex % params.print_info_interval == 0) {
|
|
|
|
|
printf("Example %d, opt iter %d\n", ex, opt->iter);
|
|
|
|
|
printf("error_before_opt: %.6f\n", error_before_opt);
|
|
|
|
|
printf("error_after_opt: %.6f\n", error_after_opt);
|
|
|
|
|
printf("used_mem_before_opt: %zu bytes\n", used_mem_before_opt);
|
|
|
|
|
printf("used_mem_after_opt: %zu bytes\n", used_mem_after_opt);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (params.print_details_interval > 0 && ex % params.print_details_interval == 0) {
|
|
|
|
|
// set_logits_masked(logits, token_notavail, -1e9);
|
|
|
|
|
for (int i=0; i<n_batch; ++i) {
|
|
|
|
|
init_sampler(&sampler, lctx);
|
|
|
|
|
for (int k=0; k<n_tokens; ++k) {
|
|
|
|
|
int32_t token = sample(&sampler,
|
|
|
|
|
(float *) ((char *) logits->data + i*logits->nb[2] + k*logits->nb[1]),
|
|
|
|
|
(llama_token *) ((char *) tokens_input->data + i*tokens_input->nb[1]),
|
|
|
|
|
k);
|
|
|
|
|
* ((int32_t *) ((char *) after_opt_best_samples->data + i*after_opt_best_samples->nb[1] + k*after_opt_best_samples->nb[0])) = token;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// printf("probabilities after optimization:\n");
|
|
|
|
|
// print_matrix(after_opt_probs);
|
|
|
|
|
printf("Example:\n---\n");
|
|
|
|
|
print_tokens_batch(lctx, tokens_input);
|
|
|
|
|
printf("\n---\n");
|
|
|
|
|
|
|
|
|
|
// printf("best samples after optimization:\n---\n");
|
|
|
|
|
printf("samples after optimization:\n---\n");
|
|
|
|
|
print_tokens_batch(lctx, after_opt_best_samples);
|
|
|
|
|
printf("\n---\n");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ggml_free(ctx0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (params.n_examples > 0) {
|
|
|
|
|
save_checkpoint(&model, opt, params.fn_checkpoint_out);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (strlen(params.fn_model_out) > 0) {
|
|
|
|
|
save_as_llama_model(&vocab, &model, params.fn_model_out);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
{
|
|
|
|
|
int n_gen = params.n_predict;
|
|
|
|
|
int sample_ctx = n_tokens - n_tokens/8;
|
|
|
|
|
|
2023-06-16 20:23:53 +02:00
|
|
|
sampler.params.temp = 0.2f;
|
|
|
|
|
sampler.params.repeat_penalty = 1.1f;
|
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
|
|
|
sampler.params.mirostat = 2;
|
|
|
|
|
init_sampler(&sampler, lctx);
|
|
|
|
|
|
|
|
|
|
printf("Generating %d tokens.\n", n_gen);
|
|
|
|
|
|
|
|
|
|
struct ggml_tensor * tokens_input = ggml_new_tensor_1d(model.ctx, GGML_TYPE_I32, n_tokens);
|
|
|
|
|
struct ggml_tensor * target_logits = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens);
|
|
|
|
|
struct ggml_tensor * target_probs = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens);
|
|
|
|
|
|
2023-08-21 22:07:43 +02:00
|
|
|
get_example_targets(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), rand()%train_samples.size(), tokens_input, target_logits, 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
|
|
|
for (int i=sample_ctx; i<n_tokens; ++i) {
|
|
|
|
|
ggml_set_i32_1d(tokens_input, i, n_vocab/2);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (int i=0; i<sample_ctx-1; ++i) {
|
|
|
|
|
print_token(lctx, ggml_get_i32_1d(tokens_input, i));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
printf("---\n");
|
|
|
|
|
for (int i=0; i<n_gen; ++i) {
|
|
|
|
|
struct ggml_init_params cparams = {
|
|
|
|
|
/*.mem_size =*/ compute_size,
|
|
|
|
|
/*.mem_buffer =*/ compute_addr,
|
|
|
|
|
/*.no_alloc =*/ false,
|
|
|
|
|
};
|
|
|
|
|
struct ggml_context * ctx0 = ggml_init(cparams);
|
|
|
|
|
|
|
|
|
|
ggml_cgraph gf = {};
|
|
|
|
|
|
|
|
|
|
int n_past = 0;
|
|
|
|
|
struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
|
|
|
|
|
|
|
|
|
|
ggml_build_forward_expand(&gf, logits);
|
2023-07-07 18:24:01 +02:00
|
|
|
ggml_graph_compute_helper(work_buffer, &gf, params.n_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
|
|
|
|
|
|
|
|
//struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
|
|
|
|
|
//struct ggml_tensor * probs = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);
|
|
|
|
|
|
|
|
|
|
// set_logits_masked(logits, token_notavail, -1e9);
|
|
|
|
|
int token = sample(&sampler,
|
|
|
|
|
(float *) ((char *) logits->data + (sample_ctx-1)*logits->nb[1]),
|
|
|
|
|
(llama_token *) tokens_input->data,
|
|
|
|
|
sample_ctx-1);
|
|
|
|
|
//int token = ggml_get_i32_1d(best_samples, sample_ctx-1);
|
|
|
|
|
|
|
|
|
|
// print_row(probs, sample_at);
|
|
|
|
|
print_token(lctx, token);
|
|
|
|
|
|
|
|
|
|
lshift_examples(tokens_input, target_logits, target_probs, 1);
|
|
|
|
|
ggml_set_i32_1d(tokens_input, 0, 0);
|
|
|
|
|
ggml_set_i32_1d(tokens_input, sample_ctx-1, token);
|
|
|
|
|
|
|
|
|
|
ggml_free(ctx0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
delete[] compute_addr;
|
|
|
|
|
delete[] compute_buf_0;
|
|
|
|
|
delete[] compute_buf_1;
|
2023-07-07 18:24:01 +02:00
|
|
|
|
2023-06-24 10:47:58 +02:00
|
|
|
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
|
|
|
ggml_free(model.ctx);
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
