haraldwolff/llama.cpp
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

train : mem usage and other improvements (#2439)

Browse source 
* fix track_max_mem in forward_batch_wo_cache_flash_attn_train

* remove unnecessary Adam(W) optimizer tensors.

reduces optimizer memory overhead from 7*modelsize to 2*modelsize.

additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t.

bumps training checkpoint file version, but old checkpoints can still be read.
new version with less tensors is saved.

* add gradient clipping to AdamW

* Fix reset of unused g->nodes and g->grads to NULL

* implement gradient checkpointing for training

reduces memory overhead from O(n_layer) to O(sqrt(n_layer))

as explained in readme of https://github.com/cybertronai/gradient-checkpointing

* remove unused compute buffer 3

* add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes

GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);

* change AdamW decay parameter to work like the torch AdamW decay parameter

It is now relative to Adam learning rate `alpha*sched`.
Before that it was relative to `sched` only.

`alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1]

* change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT

* change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW

btw: the default weight decay parameter for torch.optim.AdamW is 0.01

* bug fixes for cross entropy loss

ggml_cross_entropy_loss: sums where not correctly added in workload of each thread
ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues

guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16

cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup.
so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance.

* fix test-grad0 for cross_entropy_loss

the second argument to cross_entropy_loss must sum up to 1 for each row

* fix test-grad0 for soft_max

dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)

* improve finite differences of test-grad0 by using double instead of float

* change cross_entropy_loss to output average over all rows

this helps keeping the loss and gradients in a sane range

* improve gradient checkpointing

sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal.
since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different:

```
  given: n, u, v
  objective: minimize(a*u+b*v) where a*b=n, a>0, b>0
  b=n/a
  minimize(a*u+v*n/a)
  diff(a*u+v*n/a, a) = u - (v*n/a)/a
  diff(a*u+v*n/a, a) == 0
  u - (v*n/a)/a == 0
  u == v*n/(a*a)
  u*a*a = v*n
  a*a = v*n/u
  a = sqrt(n*v/u)
```

this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage.

* disable gradient checkpointing debug output

* llama : fix rope usage in train-text-from-scratch after ChatGLM change

* add more training parameters:

--enable-restart N         Only for Adam optimizer. Enable restarts of cos-decay
--disable-restart N        Only for Adam optimizer. Disable restarts of cos-decay
--opt-past N               Number of optimization iterations to track for delta convergence test. Disabled when zero.
--opt-delta N              Maximum delta for delta convergence test. Disabled when <= zero.
--opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero.
--adam-epsf N              AdamW epsilon for convergence test. Disabled when <= zero.
--adam-min-alpha N         Adam minimum learning rate alpha, usually 0.1 * alpha

* replace memcpy with reshape operation so that the graph is not cut at the input

this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it

* remove unused function argument from get_example_targets_batch

* measure and print total training time

* add optimization callback to ggml_opt_resume_g

this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)).

can be used for dynamic learning schedule and setting input data for batches before each iteration

* use optimization callback in training

allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters

reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration

* add minimum number of tensor dimensions to apply weight decay (default 2)

this allows to not apply weight decay to bias parameters

* rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup

* fix increase of model.train_samples and model.train_tokens

now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations

* change sampling parameters for prediction after training to defaults of common.h

and clarify what is context for prediction and what are generated tokens

* tighten abs error bounds for cross_entropy_loss in test-grad0

* add conditional compilation of using F16 exp in flash attention

uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention

* tighten abs error bounds for flash_attn in test-grad0

* tighten abs error bounds for sqrt in test-grad0

* remove out-commented vectorized code of opt_adam

the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead

* ggml : update ggml_rms_norm_back with configurable eps

* llama training : fix ggml_rms_norm_back calls to pass configurable eps

* remove trailing whitespace

* add train function using automatic gradient checkpointing backward pass and allocator

* in train function replace add_inplace by regular add

because using add_inplace seems to result in different gradients

* don't use allocate hash_map on context

because the context has no_alloc=True when using memory allocator resulting in NULL data pointers

* correctly clone reshape and permute operations by also cloning tensor->nb values

* fix variable name and add missing type cast

* terminate recursive tensor cloning when reaching tensor without src tensors

* correctly clone view tensors by setting data pointers

without this the checkpointing would only work when being used together with memory allocator

* fix variable names

* swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn`

* add input tensors as checkpoints

so that recursive tensor cloning of gradient checkpointing terminates on input tensors

* fix variable name and add missing boolean negation

* make sure some tensors are not reallocated by inserting new temporary nodes depending on them:

output and parameter gradient tensors need to be available at the end of the graph execution

parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration

checkpoint tensors are allocated all together to reduce memory allocator fragmentation

afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs

* fix ASSERT to work with zero layers

* add training options whether to use allocator and/or unified training function

* integrate unified training function which may use memory allocator

the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing

* format name of cloned tensors with " (clone)" suffix

* set names for tensors in unified train function for easier debugging

* allocate graph on context using ggml_new_graph

* remove handwritten training functions

* remove unused training parameters "use_scratch" and "use_unified"

* remove trailing whitespace

* remove unused train params: mem_compute1_gb & mem_compute2_gb

mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)

* remove unused forward_batch function

* add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly

* only use ggml_allocr_alloc when tensor has NULL data and is no view

* fix test when to create temporary backward graph

temporary backward graph is only necessary when using checkpointing

* fix memory "leak" in optimizers

each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.

* reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator

with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.

the computation results are the same

* add missing lctx argument to get_example_targets_batch

* implement llama model file saving using gguf

checkpoint loading and saving disabled, to be replaced by loading and saving via gguf

* implement loading/saving of checkpointing files using GGUF

* bug fixes

* add checkpoint file version for future compatibility

* update readme with gguf filenames

* save & load opt->just_initialized value

* add first draft for checkpoint conversion script

* add gguf arch and ftype

* save opt parameter counter as uint64

* add gguf key and tensor names for optimizer and training

* add layer_norm_rms_eps to checkpoint convert script

* use same GGUF_GET_KEY macro as in llama.cpp

* use norm_rms_eps, and rope parameters and command line options to set them

* fix memory corruption bug in gguf

ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free.
to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function.
so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying
and freeing the old data.

* add gguf example cmake file

* bug fixes in tokenize_file

* bug fixes in load_llama_model_gguf

* bug fix: init model when no checkpoint was loaded

* bug fix in read_tensor_by_name

* bug fix in load_opt_context_gguf

* avoid printing lots of spaced on the unusual case that loss gets nan

* set name of tensors with empty name from what was read from gguf

* remove trailing whitespace

* print data checksums before saving and after loading to verify correctness

* bug fixes for convert-train-checkpoint-to-gguf

* temporarily add code to write old checkpoint files

used to verify that old checkpoint files are correctly converted to gguf

* bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0

* remove code used to verify correctness of checkpoint file conversion

* remove trailing whitespace

* remove prediction related code

use main for prediction, it is better optimized

* update train-text-from-scratch README.md

* fix non-windows GGML_ALIGNED_REALLOC

* add missing blank line at end of file

* remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos

* train : fix compile warnings

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
xaedes 2023-08-28 21:51:47 +02:00 committed by GitHub
parent 43033b7bb4
commit 44c117f41e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
11 changed files with 1940 additions and 2508 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
common/common.cpp
View file

@ -15,6 +15,7 @@
#include <string>
#include <unordered_set>
#include <vector>
#include <cinttypes>
#if defined(__APPLE__) && defined(__MACH__)
#include <sys/types.h>
@ -938,8 +939,8 @@ std::string get_sortable_timestamp() {
const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
current_time.time_since_epoch() % 1000000000).count();
char timestamp_ns[10];
snprintf(timestamp_ns, 11, "%09ld", ns);
char timestamp_ns[11];
snprintf(timestamp_ns, 11, "%09" PRId64, ns);
return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
}

1
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -681,7 +681,6 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
// for rms-att-weight
int row_length = model->hparams.n_embd;
const auto & hparams = model->hparams;
int n_ff = model->hparams.n_ff;
for (uint32_t i = 0; i < model->hparams.n_layer; ++i){

5
examples/gguf/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET gguf)
add_executable(${TARGET} gguf.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

14
examples/train-text-from-scratch/README.md
View file

@ -8,15 +8,15 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
# train
./bin/train-text-from-scratch \
--vocab-model ../models/ggml-vocab.bin \
--vocab-model ../models/ggml-vocab-llama.gguf \
--ctx 64 --embd 256 --head 8 --layer 16 \
--checkpoint-in chk-shakespeare-256x16.bin \
--checkpoint-out chk-shakespeare-256x16.bin \
--model-out ggml-shakespeare-256x16-f32.bin \
--checkpoint-in chk-shakespeare-256x16.gguf \
--checkpoint-out chk-shakespeare-256x16.gguf \
--model-out ggml-shakespeare-256x16-f32.gguf \
--train-data "shakespeare.txt" \
-t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
--print-details-interval 0 --predict 16 --use-flash
-t 6 -b 16 --seed 1 --adam-iter 256 \
--no-checkpointing
# predict
./bin/main -m ggml-shakespeare-256x16-f32.bin
./bin/main -m ggml-shakespeare-256x16-f32.gguf
```

492
examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py Normal file
View file

@ -0,0 +1,492 @@
#!/usr/bin/env python3
# train-text-from-scratch checkpoint --> gguf conversion
import argparse
import gguf
import os
import struct
import sys
import numpy as np
from pathlib import Path
# gguf constants
LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
LLM_KV_OPTIMIZER_TYPE_ADAM = "adam"
LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version"
LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count"
LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count"
LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count"
LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized"
LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss"
LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss"
LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count"
LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss"
LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step"
LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j"
LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k"
LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end"
LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments"
LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments"
LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters"
LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients"
LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients"
LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction"
LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values"
LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha"
LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys"
LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s"
LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y"
LLM_KV_TRAINING_FILE_VERSION = "training.file_version"
LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count"
LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count"
LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count"
class Tensor:
def __init__(self, dtype='f', ne=None):
if ne is None:
ne = []
self.dtype = dtype
self.ne = ne
self.nbytes = 0
if self.dtype == 'f':
if len(self.ne) == 0:
self.nbytes = 0
else:
self.nbytes = int(np.product(self.ne)) * 4
else:
raise ValueError(f"Unhandled data type '{self.dtype}'")
def load(self, data, offset):
nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
assert(nd == len(self.ne))
ne = []
for d in range(nd):
n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
ne.append(n)
assert(tuple(ne) == tuple(self.ne))
if self.dtype == 'f':
assert(dtype == 0)
else:
raise ValueError(f"Unhandled data type '{self.dtype}'")
self.name = bytes(data[offset:offset+namelen]); offset += namelen
# 32-byte alignment
offset += (0 - offset) & 31
self.data = data[offset:offset+self.nbytes]
offset += self.nbytes
return offset
def max_storage_size(self):
result = 0
result += 4 # nd
result += 4 # namelen
result += 4 # dtype
result += len(self.ne)*8 # ne
result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
result += 31 # 32-byte alignment
result += self.nbytes
return result
def save_gguf(self, gguf_writer, name):
gguf_writer.add_tensor(
name=name,
tensor=self.data,
raw_shape=np.array(list(reversed(self.ne))),
raw_dtype=gguf.GGMLQuantizationType.F32)
class OptimizationParamsV0:
def __init__(self):
pass
def load(self, data, offset):
self.type = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_threads = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.past = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.delta = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.print_forward_graph = struct.unpack('<?', bytes(data[offset:offset + 1]))[0]; offset += 4 # 32bit-aligned
self.print_backward_graph = struct.unpack('<?', bytes(data[offset:offset + 1]))[0]; offset += 4 # 32bit-aligned
self.adam_n_iter = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_sched = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_decay = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_alpha = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_beta1 = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_beta2 = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_eps = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_eps_f = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_eps_g = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_n_iter = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_max_linesearch = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_eps = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_ftol = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_wolfe = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_min_step = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_max_step = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_linesearch = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
return offset
class OptimizationContext:
def __init__(self):
pass
def load(self, data, offset):
self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
offset += 4
if self.version == 0:
params = OptimizationParamsV0()
offset = params.load(data, offset)
self.past = params.past
self.lbfgs_m = params.lbfgs_m
self.nx = struct.unpack('N', bytes(data[offset:offset + 8]))[0]; offset += 8
self.iter = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]); offset += 4
self.type = params.type
self.adam_m = Tensor('f', [self.nx])
self.adam_v = Tensor('f', [self.nx])
self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
self.lbfgs_x = Tensor('f', [self.nx])
self.lbfgs_xp = Tensor('f', [self.nx])
self.lbfgs_g = Tensor('f', [self.nx])
self.lbfgs_gp = Tensor('f', [self.nx])
self.lbfgs_d = Tensor('f', [self.nx])
self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else [])
self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m])
self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m])
if self.type == 0:
# these tensors are stored, but we don't need their data
x = Tensor('f', [self.nx])
g = Tensor('f', [self.nx])
g2 = Tensor('f', [self.nx])
mh = Tensor('f', [self.nx])
vh = Tensor('f', [self.nx])
offset = x.load(data, offset)
offset = g.load(data, offset)
offset = g2.load(data, offset)
offset = self.adam_m.load(data, offset)
offset = self.adam_v.load(data, offset)
offset = mh.load(data, offset)
offset = vh.load(data, offset)
offset = self.adam_pf.load(data, offset)
self.adam_fx_best = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_fx_prev = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
elif self.type == 1:
offset = self.lbfgs_x.load(data, offset)
offset = self.lbfgs_xp.load(data, offset)
offset = self.lbfgs_g.load(data, offset)
offset = self.lbfgs_gp.load(data, offset)
offset = self.lbfgs_d.load(data, offset)
offset = self.lbfgs_pf.load(data, offset)
offset = self.lbfgs_lmal.load(data, offset)
offset = self.lbfgs_lmys.load(data, offset)
offset = self.lbfgs_lms.load(data, offset)
offset = self.lbfgs_lmy.load(data, offset)
self.lbfgs_fx_best = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_step = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_j = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_k = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_end = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
else:
raise ValueError('Unknown optimizer type')
elif self.version == 1:
self.past = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.nx = struct.unpack('N', bytes(data[offset:offset + 8]))[0]; offset += 8
self.iter = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]); offset += 4
self.adam_m = Tensor('f', [self.nx])
self.adam_v = Tensor('f', [self.nx])
self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
self.lbfgs_x = Tensor('f', [self.nx])
self.lbfgs_xp = Tensor('f', [self.nx])
self.lbfgs_g = Tensor('f', [self.nx])
self.lbfgs_gp = Tensor('f', [self.nx])
self.lbfgs_d = Tensor('f', [self.nx])
self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else [])
self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m])
self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m])
# forgot to save type in version 1:
# guess self.type from number of remaining bytes
size_type_0 = 12 + sum([t.max_storage_size() for t in
[self.adam_m, self.adam_v]
+([self.adam_pf] if (self.past > 0) else [])])
size_type_1 = 24 + sum([t.max_storage_size() for t in
[self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
self.lbfgs_lmal, self.lbfgs_lmys,
self.lbfgs_lms, self.lbfgs_lmy]
+([self.lbfgs_pf] if (self.past > 0) else [])])
# due to alignment padding the size might not by exact
# but the difference in size for both types is significant,
# so we can just use whichever is closest
remaining = len(data) - offset
if abs(remaining - size_type_0) < abs(remaining - size_type_1):
self.type = 0
else:
self.type = 1
if self.type == 0:
offset = self.adam_m.load(data, offset)
offset = self.adam_v.load(data, offset)
offset = self.adam_pf.load(data,offset)
self.adam_fx_best = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_fx_prev = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
elif self.type == 1:
offset = self.lbfgs_x.load(data, offset)
offset = self.lbfgs_xp.load(data, offset)
offset = self.lbfgs_g.load(data, offset)
offset = self.lbfgs_gp.load(data, offset)
offset = self.lbfgs_d.load(data, offset)
offset = self.lbfgs_pf.load(data, offset)
offset = self.lbfgs_lmal.load(data, offset)
offset = self.lbfgs_lmys.load(data, offset)
offset = self.lbfgs_lms.load(data, offset)
offset = self.lbfgs_lmy.load(data, offset)
self.lbfgs_fx_best = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_step = struct.unpack('<f', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_j = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_k = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_end = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0]; offset += 4
else:
raise ValueError('Invalid version of checkpoint file')
return offset
def save_gguf(self, gguf_writer):
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
if self.type == 0:
gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
if self.past > 0:
self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
elif self.type == 1:
gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
if self.past > 0:
self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
else:
raise ValueError('Unknown optimizer type')
class ModelParams:
def __init__(self):
pass
def load(self, data, offset):
self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_embd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_mult = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_head = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.n_rot = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
return offset
def get_n_ff(self):
# struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
def save_gguf(self, gguf_writer):
# self.n_vocab not saved
gguf_writer.add_embedding_length(self.n_embd)
gguf_writer.add_head_count(self.n_head)
gguf_writer.add_block_count(self.n_layer)
gguf_writer.add_rope_dimension_count(self.n_rot)
gguf_writer.add_feed_forward_length(self.get_n_ff())
def tensor_name(key, bid=None):
return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + ".weight"
class Layer:
def __init__(self, params, bid):
self.bid = bid
self.att_norm = Tensor('f', [params.n_embd])
self.wq = Tensor('f', [params.n_embd, params.n_embd])
self.wk = Tensor('f', [params.n_embd, params.n_embd])
self.wv = Tensor('f', [params.n_embd, params.n_embd])
self.wo = Tensor('f', [params.n_embd, params.n_embd])
self.ffn_norm = Tensor('f', [params.n_embd])
self.w1 = Tensor('f', [params.n_embd, params.get_n_ff()])
self.w2 = Tensor('f', [params.get_n_ff(), params.n_embd])
self.w3 = Tensor('f', [params.n_embd, params.get_n_ff()])
def load(self, data, offset):
offset = self.att_norm.load(data, offset)
offset = self.wq.load(data, offset)
offset = self.wk.load(data, offset)
offset = self.wv.load(data, offset)
offset = self.wo.load(data, offset)
offset = self.ffn_norm.load(data, offset)
offset = self.w1.load(data, offset)
offset = self.w2.load(data, offset)
offset = self.w3.load(data, offset)
return offset
def save_gguf(self, gguf_writer):
self.att_norm.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid))
self.wq.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q, self.bid))
self.wk.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K, self.bid))
self.wv.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V, self.bid))
self.wo.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, self.bid))
self.ffn_norm.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM, self.bid))
self.w1.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE, self.bid))
self.w2.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, self.bid))
self.w3.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP, self.bid))
class Model:
def __init__(self):
self.params = ModelParams()
self.layers = []
def load(self, data, offset):
offset = self.params.load(data, offset)
self.tok_embd = Tensor('f', [self.params.n_embd, self.params.n_vocab])
self.norm = Tensor('f', [self.params.n_embd])
self.output = Tensor('f', [self.params.n_embd, self.params.n_vocab])
offset = self.tok_embd.load(data, offset)
offset = self.norm.load(data, offset)
offset = self.output.load(data, offset)
self.layers.clear()
for bid in range(self.params.n_layer):
layer = Layer(self.params, bid)
offset = layer.load(data, offset)
self.layers.append(layer)
return offset
def save_gguf(self, gguf_writer):
self.params.save_gguf(gguf_writer)
self.tok_embd.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD))
self.norm.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM))
self.output.save_gguf (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT))
for layer in self.layers:
layer.save_gguf(gguf_writer)
class Checkpoint:
def __init__(self):
self.model = Model()
self.opt_ctx = OptimizationContext()
def load(self, data, offset):
magic = bytes(reversed(data[offset:offset + 4])); offset += 4
if magic != b'ggcp':
raise ValueError(f"File header magic indicates, that this is no checkpoint file. Expected 'ggcp', Got '{str(magic)}'")
self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
if self.version != 0:
raise ValueError('Invalid version of checkpoint file')
self.train_its = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
self.train_tokens = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
offset = self.model.load(data, offset)
offset = self.opt_ctx.load(data, offset)
return offset
def save_gguf(self, gguf_writer):
gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
gguf_writer.add_layer_norm_rms_eps(1e-5)
gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION, 0)
gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT, self.train_samples)
gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT, self.train_tokens)
self.model.save_gguf(gguf_writer)
self.opt_ctx.save_gguf(gguf_writer)
def handle_args():
parser = argparse.ArgumentParser(description = 'Convert train-text-from-scratch checkpoints to GGUF')
parser.add_argument('--input', '-i', type = Path, help = 'Input train checkpoint filename', required=True)
parser.add_argument('--output', '-o', type = Path, help ='Output GGUF filename', required=True)
return parser.parse_args()
def main():
cfg = handle_args()
data = np.memmap(cfg.input, mode = 'r')
chk = Checkpoint()
offset = 0
offset = chk.load(data, offset)
# we should have read all available data
assert(offset == len(data))
gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
chk.save_gguf(gguf_writer)
print(" gguf: write header")
gguf_writer.write_header_to_file()
print(" gguf: write metadata")
gguf_writer.write_kv_data_to_file()
print(" gguf: write tensors")
gguf_writer.write_tensors_to_file()
gguf_writer.close()
if __name__ == '__main__':
main()

3502
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

File diff suppressed because it is too large Load diff

4
ggml-alloc.c
View file

@ -107,6 +107,10 @@ static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct g
}
void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
#ifdef GGML_ALLOCATOR_DEBUG
GGML_ASSERT(ggml_is_view(tensor) == false); // views generally get data pointer from one of their sources
GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
#endif
size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, alloc->alignment);

335
ggml.c
View file

@ -123,6 +123,8 @@ typedef void * thread_ret_t;
#define GGML_GELU_FP16
#define GGML_GELU_QUICK_FP16
#define GGML_SILU_FP16
// #define GGML_CROSS_ENTROPY_EXP_FP16
// #define GGML_FLASH_ATTN_EXP_FP16
#define GGML_SOFT_MAX_UNROLL 4
#define GGML_VEC_DOT_UNROLL 2
@ -186,8 +188,8 @@ typedef void * thread_ret_t;
//
#if defined(_MSC_VER) || defined(__MINGW32__)
#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
#else
inline static void * ggml_aligned_malloc(size_t size) {
void * aligned_memory = NULL;
@ -212,8 +214,8 @@ inline static void * ggml_aligned_malloc(size_t size) {
}
return aligned_memory;
}
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#define GGML_ALIGNED_FREE(ptr) free(ptr)
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#define GGML_ALIGNED_FREE(ptr) free(ptr)
#endif
#define UNUSED GGML_UNUSED
@ -5857,7 +5859,8 @@ struct ggml_tensor * ggml_rms_norm_inplace(
struct ggml_tensor * ggml_rms_norm_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b) {
struct ggml_tensor * b,
float eps) {
bool is_node = false;
if (a->grad) {
@ -5867,6 +5870,8 @@ struct ggml_tensor * ggml_rms_norm_back(
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, &eps, sizeof(eps));
result->op = GGML_OP_RMS_NORM_BACK;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
@ -9443,6 +9448,8 @@ static void ggml_compute_forward_div_f32(
#ifdef GGML_USE_ACCELERATE
UNUSED(ggml_vec_div_f32);
vDSP_vdiv(
(float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
(float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
@ -10749,7 +10756,8 @@ static void ggml_compute_forward_rms_norm_back_f32(
GGML_TENSOR_BINARY_OP_LOCALS;
const float eps = 1e-6f; // TODO: make this a parameter
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
// TODO: optimize
for (int64_t i03 = 0; i03 < ne03; i03++) {
@ -12139,6 +12147,7 @@ static void ggml_compute_forward_soft_max_back_f32(
// dx = J * dy
// dxk = sum_i(Jki * dyi)
// dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk
// dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk
// dxk = sum_i(-yk*yi * dyi) + yk*dyk
// dxk = -yk * sum_i(yi * dyi) + yk*dyk
// dxk = -yk * dot(y, dy) + yk*dyk
@ -13929,7 +13938,7 @@ static void ggml_compute_forward_flash_attn_f32(
vvexpf(S, S, &Mup);
ggml_vec_sum_f32(Mup, &sum, S);
#else
uint16_t scvt[GGML_SOFT_MAX_UNROLL];
uint16_t scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
@ -13939,9 +13948,13 @@ static void ggml_compute_forward_flash_attn_f32(
if (SS[j] == -INFINITY) {
SS[j] = 0.0f;
} else {
#ifndef GGML_FLASH_ATTN_EXP_FP16
const float val = expf(SS[j] - max);
#else
ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
memcpy(&scvt[j], &s, sizeof(uint16_t));
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
#endif
sump[j] += (ggml_float)val;
SS[j] = val;
}
@ -14519,7 +14532,7 @@ static void ggml_compute_forward_flash_attn_back_f32(
vvexpf(SM, SM, &Mup);
ggml_vec_sum_f32(Mup, &sum, SM);
#else
uint16_t scvt[GGML_SOFT_MAX_UNROLL];
uint16_t scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
@ -14530,9 +14543,13 @@ static void ggml_compute_forward_flash_attn_back_f32(
if (SR[j] == -INFINITY) {
SW[j] = 0.0f;
} else {
#ifndef GGML_FLASH_ATTN_EXP_FP16
const float val = expf(SR[j] - max);
#else
ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max);
memcpy(&scvt[j], &s, sizeof(uint16_t));
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
#endif
sump[j] += (ggml_float)val;
SW[j] = val;
}
@ -15270,6 +15287,8 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
const int nc = src0->ne[0];
const int nr = ggml_nrows(src0);
GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc));
if (params->type == GGML_TASK_INIT) {
if (ith == 0) {
memset(sums, 0, sizeof(float) * (nth + nth * nc));
@ -15281,7 +15300,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
if (ith == 0) {
float * dp = (float *) dst->data;
ggml_vec_sum_f32(nth, dp, sums);
dp[0] *= -1.0f;
dp[0] *= -1.0f / (float) nr;
}
return;
}
@ -15298,7 +15317,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
for (int i1 = ir0; i1 < ir1; i1++) {
float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]);
float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]);
float * st = (float *) params->wdata + nth + ith*nc;
float * st = ((float *) params->wdata) + nth + ith*nc;
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
@ -15313,15 +15332,19 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
float max = -INFINITY;
ggml_vec_max_f32(nc, &max, s0);
uint16_t scvt;
uint16_t scvt; UNUSED(scvt);
for (int i = 0; i < nc; i++) {
if (s0[i] == -INFINITY) {
st[i] = 0.0f;
} else {
// const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max);
#ifndef GGML_CROSS_ENTROPY_EXP_FP16
const float s = s0[i] - max;
const float val = expf(s);
#else
ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
memcpy(&scvt, &s, sizeof(scvt));
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
#endif
sum += (ggml_float)val;
st[i] = val;
}
@ -15337,7 +15360,9 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
ggml_vec_log_f32(nc, st, st);
ggml_vec_mul_f32(nc, st, st, s1);
ggml_vec_sum_f32(nc, sums + ith, st);
float st_sum = 0;
ggml_vec_sum_f32(nc, &st_sum, st);
sums[ith] += st_sum;
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
@ -15387,7 +15412,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
return;
}
const float eps = 1e-9f;
const double eps = 1e-9;
// TODO: handle transposed/permuted matrices
const int64_t nc = src0->ne[0];
@ -15406,7 +15431,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
float * ds0 = (float *)((char *) dst->data + i1*dst->nb[1]);
float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]);
float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]);
float * sm = (float *) params->wdata + ith*nc;
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
@ -15415,54 +15439,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
assert(!isnan(s1[i]));
}
#endif
// step by step explanation:
{
//float * sums = (float *) params->wdata;
// forward pass with annotated gradients from backward pass
// (built by going in reverse operation order, adding to gradients of current operation args)
// st0 = exp(s0-max(s0)) grad[st0] = grad[st1]*(1.0 - eps)/sum
// from softmax_back: grad[s0] = st1_k * (grad[st1]_k - dot(st1, grad[st1]))
// ggml_vec_scale_f32(nc, st, sum); // st1 = st0*/sum = softmax(s0) grad[st1] = grad[st2]*(1.0 - eps)
// ggml_vec_scale_f32(nc, st, (1.0f - eps)); // st2 = st1*(1.0 - eps) grad[st2] = grad[st3]
// ggml_vec_add1_f32(nc, st, st, eps); // st3 = st2 + eps grad[st3] = grad[st4]/st3
// ggml_vec_log_f32(nc, st, st); // st4 = log(st3) grad[st4] = grad[st5] * s1
// ggml_vec_mul_f32(nc, st, st, s1); // st5 = st4 * s1 grad[st5] = grad[sums[ith]]
// ggml_vec_sum_f32(nc, sums + ith, st); // sums[ith] = st5 grad[sums[ith]] = grad[cross_entropy_loss] = -grad[cel]
// substitute into grad[st1], because we can reuse softmax_back from this point on
// grad[st1] = -grad[cel]*s1*(1.0 - eps)/(eps + softmax(s0)*(1.0 - eps))
// postorder:
// grad[st1] := softmax(s0)
// grad[st1] := grad[st1]*(1.0 - eps)
// grad[st1] := grad[st1] + eps
// grad[st1] := s1 / grad[st1]
// grad[st1] := grad[st1]*(1.0-eps)*-grad[cel]
// src0 gradients by going through softmax_back
// grad[s0] = st1_k * (grad[st1]_k - dot(st1, grad[st1]))
// from softmax_back:
// dxk = yk * (dyk - dot(y, dy))
// dot_y_dy := dot(y, dy)
// dx := dy
// dx := dx - dot_y_dy
// dx := dx * y
// postorder:
// dot_st1_dst1 := dot(st1, grad[st1])
// grad[s0] := grad[st1]
// grad[s0] := grad[s0] - dot_st1_dst1
// grad[s0] := grad[s0] * st1
// prepend postorder from grad[st1] directly using grad[s0] as memory location, as we will grad[s0] := grad[st1]
// sm := softmax(s0)
// grad[s0] := sm*(1.0 - eps)
// grad[s0] := grad[s0] + eps
// grad[s0] := s1 / grad[s0]
// grad[s0] := grad[s0]*(1.0-eps)*-grad[cel]
// dot_st1_dst1 := dot(sm, grad[s0])
// grad[s0] := grad[s0] - dot_st1_dst1
// grad[s0] := grad[s0] * sm
}
// soft_max
ggml_float sum = 0.0;
@ -15470,39 +15446,37 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
float max = -INFINITY;
ggml_vec_max_f32(nc, &max, s0);
uint16_t scvt;
uint16_t scvt; UNUSED(scvt);
for (int i = 0; i < nc; i++) {
if (s0[i] == -INFINITY) {
sm[i] = 0.0f;
ds0[i] = 0.0f;
} else {
// const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max);
#ifndef GGML_CROSS_ENTROPY_EXP_FP16
const float s = s0[i] - max;
const float val = expf(s);
#else
ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
memcpy(&scvt, &s, sizeof(scvt));
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
#endif
sum += (ggml_float)val;
sm[i] = val;
ds0[i] = val;
}
}
assert(sum > 0.0);
sum = 1.0/sum;
sum = (1.0 - eps)/sum;
}
float dot_st1_dst1 = 0;
ggml_vec_scale_f32(nc, sm, sum);
ggml_vec_cpy_f32 (nc, ds0, sm);
ggml_vec_scale_f32(nc, ds0, (1.0f - eps));
ggml_vec_add1_f32 (nc, ds0, ds0, eps);
ggml_vec_div_f32 (nc, ds0, s1, ds0);
ggml_vec_scale_f32(nc, ds0, -(1.0f - eps)*d[0]);
ggml_vec_dot_f32 (nc, &dot_st1_dst1, sm, ds0);
ggml_vec_acc1_f32 (nc, ds0, -dot_st1_dst1);
ggml_vec_mul_f32 (nc, ds0, ds0, sm);
// grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr
ggml_vec_scale_f32(nc, ds0, sum);
ggml_vec_add1_f32(nc, ds0, ds0, eps);
ggml_vec_sub_f32(nc, ds0, ds0, s1);
ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr);
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
assert(!isnan(sm[i]));
assert(!isinf(sm[i]));
assert(!isnan(ds0[i]));
assert(!isinf(ds0[i]));
}
@ -16057,9 +16031,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
{
// necessary for llama
if (src0->grad) {
float eps;
memcpy(&eps, tensor->op_params, sizeof(float));
src0->grad = ggml_add_impl(ctx,
src0->grad,
ggml_rms_norm_back(ctx, src0, tensor->grad),
ggml_rms_norm_back(ctx, src0, tensor->grad, eps),
inplace);
}
} break;
@ -16827,9 +16804,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) {
return result;
}
struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) {
struct ggml_cgraph result = *gf;
void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) {
GGML_ASSERT(gf->n_nodes > 0);
// if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph
@ -16853,15 +16828,19 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
}
}
for (int i = gf->n_nodes - 1; i >= 0; i--) {
for (int i = 0; i < gf->n_nodes; i++) {
struct ggml_tensor * node = gf->nodes[i];
if (node->is_param) {
GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
ggml_build_forward_expand(&result, node->grad);
ggml_build_forward_expand(gb, node->grad);
}
}
}
struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) {
struct ggml_cgraph result = *gf;
ggml_build_backward_expand(ctx, gf, &result, keep);
return result;
}
@ -17537,10 +17516,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
{
n_tasks = n_threads;
size_t cur = ggml_type_size(node->type)*node->src[0]->ne[0]*n_tasks;
work_size = MAX(work_size, cur);
} break;
case GGML_OP_NONE:
{
@ -18418,14 +18393,16 @@ static enum ggml_opt_result ggml_opt_adam(
struct ggml_opt_params params,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb) {
struct ggml_cgraph * gb,
ggml_opt_callback callback,
void * callback_data) {
GGML_ASSERT(ggml_is_scalar(f));
// these will store the parameters we want to optimize
struct ggml_tensor * ps[GGML_MAX_PARAMS];
int np = 0;
int nx = 0;
int64_t nx = 0;
for (int i = 0; i < gf->n_nodes; ++i) {
if (gf->nodes[i]->is_param) {
GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
@ -18444,31 +18421,32 @@ static enum ggml_opt_result ggml_opt_adam(
}
// constants
const float sched = params.adam.sched;
const float decay = params.adam.decay * sched;
const float alpha = params.adam.alpha * sched;
float sched = params.adam.sched;
const float alpha = params.adam.alpha;
const float decay = params.adam.decay * alpha;
const float beta1 = params.adam.beta1;
const float beta2 = params.adam.beta2;
const float eps = params.adam.eps;
const float gclip = params.adam.gclip;
const int decay_min_ndim = params.adam.decay_min_ndim;
float * x = opt->adam.x->data; // view of the parameters
float * g1 = opt->adam.g1->data; // gradient
float * g2 = opt->adam.g2->data; // gradient squared
float * m = opt->adam.m->data; // first moment
float * v = opt->adam.v->data; // second moment
float * mh = opt->adam.mh->data; // first moment hat
float * vh = opt->adam.vh->data; // second moment hat
float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values
// update view
ggml_opt_get_params(np, ps, x);
if (callback) {
callback(callback_data, &sched);
}
// compute the function value
ggml_graph_reset (gf);
ggml_set_f32 (f->grad, 1.0f);
ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
ggml_graph_compute(gb, &cplan);
opt->adam.fx_prev = ggml_get_f32_1d(f, 0);
opt->adam.fx_best = opt->adam.fx_prev;
@ -18476,6 +18454,9 @@ static enum ggml_opt_result ggml_opt_adam(
pf[opt->iter % params.past] = opt->adam.fx_prev;
}
opt->loss_before = opt->adam.fx_prev;
opt->loss_after = opt->adam.fx_prev;
// initialize
if (opt->just_initialized) {
opt->adam.n_no_improvement = 0;
@ -18508,50 +18489,55 @@ static enum ggml_opt_result ggml_opt_adam(
UNUSED(t_start_cpu);
{
// update the gradient
ggml_opt_get_grad(np, ps, g1);
float gnorm = 1.0f;
if (gclip > 0.0f) {
// gradient clipping
ggml_float sum = 0.0;
for (int p = 0; p < np; ++p) {
const int64_t ne = ggml_nelements(ps[p]);
for (int64_t j = 0; j < ne; ++j) {
float g = ggml_get_f32_1d(ps[p]->grad, j);
sum += (ggml_float)(g*g);
}
}
ggml_float norm = sqrt(sum);
if (norm > (ggml_float) gclip) {
gnorm = (float) ((ggml_float) gclip / norm);
}
}
const float beta1h = alpha*sched/(1.0f - powf(beta1, opt->iter));
const float beta2h = 1.0f/(1.0f - powf(beta2, opt->iter));
int64_t i = 0;
for (int p = 0; p < np; ++p) {
const int64_t ne = ggml_nelements(ps[p]);
const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
for (int64_t j = 0; j < ne; ++j) {
float x = ggml_get_f32_1d(ps[p], j);
float g = ggml_get_f32_1d(ps[p]->grad, j)*gnorm;
m[i] = m[i]*beta1 + g*(1.0f - beta1);
v[i] = v[i]*beta2 + g*g*(1.0f - beta2);
float mh = m[i]*beta1h;
float vh = v[i]*beta2h;
vh = sqrtf(vh) + eps;
x = x*(1.0f - p_decay) - mh/vh;
ggml_set_f32_1d(ps[p], j, x);
++i;
}
}
}
// m_t = beta1*m_t-1 + (1 - beta1)*g_t
ggml_vec_scale_f32(nx, m, beta1);
ggml_vec_mad_f32 (nx, m, g1, 1.0f - beta1);
// g2 = g1^2
ggml_vec_sqr_f32 (nx, g2, g1);
// v_t = beta2*v_t-1 + (1 - beta2)*g_t^2
ggml_vec_scale_f32(nx, v, beta2);
ggml_vec_mad_f32 (nx, v, g2, 1.0f - beta2);
// m^hat = m_t / (1 - beta1^t)
// v^hat = v_t / (1 - beta2^t)
// x_t = x_t-1 - sched*(alpha*m^hat/(sqrt(v^hat) + eps) + decay*x_t-1)
// x_t = x_t-1 - sched*alpha*m^hat/(sqrt(v^hat) + eps) - sched*decay*x_t-1
// x_t = x_t-1*(1-sched*decay) - sched*alpha*m^hat/(sqrt(v^hat) + eps)
// x_t = x_t-1*(1-sched*decay) + sched*decay*(-alpha/decay)*m^hat/(sqrt(v^hat) + eps)
// x_t = mix(x_t-1, (-alpha/decay)*m^hat/(sqrt(v^hat) + eps), sched*decay)
ggml_vec_cpy_f32 (nx, mh, m);
ggml_vec_cpy_f32 (nx, vh, v);
ggml_vec_scale_f32(nx, mh, alpha/(1.0f - powf(beta1, opt->iter)));
ggml_vec_scale_f32(nx, vh, 1.0f/(1.0f - powf(beta2, opt->iter)));
ggml_vec_sqrt_f32 (nx, vh, vh);
ggml_vec_acc1_f32 (nx, vh, eps);
ggml_vec_div_f32 (nx, mh, mh, vh);
ggml_vec_scale_f32(nx, x, 1.0f - decay);
ggml_vec_sub_f32 (nx, x, x, mh);
// update the parameters
ggml_opt_set_params(np, ps, x);
if (callback) {
callback(callback_data, &sched);
}
ggml_graph_reset (gf);
ggml_set_f32 (f->grad, 1.0f);
ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
ggml_graph_compute(gb, &cplan);
const float fx = ggml_get_f32_1d(f, 0);
opt->loss_after = fx;
// check convergence
if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) {
@ -18620,7 +18606,6 @@ struct ggml_lbfgs_iteration_data {
};
static enum ggml_opt_result linesearch_backtracking(
struct ggml_context * ctx,
const struct ggml_opt_params * params,
int nx,
float * x,
@ -18632,8 +18617,11 @@ static enum ggml_opt_result linesearch_backtracking(
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cplan * cplan,
const int np,
struct ggml_tensor * ps[]) {
struct ggml_tensor * ps[],
ggml_opt_callback callback,
void * callback_data) {
int count = 0;
float width = 0.0f;
@ -18662,6 +18650,12 @@ static enum ggml_opt_result linesearch_backtracking(
dgtest = params->lbfgs.ftol*dginit;
while (true) {
if (callback) {
// LBFG-S does not support learning rate -> ignore learning schedule
float sched = 0;
callback(callback_data, &sched);
}
ggml_vec_cpy_f32(nx, x, xp);
ggml_vec_mad_f32(nx, x, d, *step);
@ -18672,7 +18666,7 @@ static enum ggml_opt_result linesearch_backtracking(
ggml_graph_reset (gf);
ggml_set_f32 (f->grad, 1.0f);
ggml_graph_compute_with_ctx(ctx, gb, params->n_threads);
ggml_graph_compute(gb, cplan);
ggml_opt_get_grad(np, ps, g);
@ -18732,7 +18726,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
struct ggml_opt_params params,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb) {
struct ggml_cgraph * gb,
ggml_opt_callback callback,
void * callback_data) {
if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE ||
params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) {
@ -18764,6 +18760,10 @@ static enum ggml_opt_result ggml_opt_lbfgs(
opt->iter = iter;
}
struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
float * x = opt->lbfgs.x->data; // current parameters
float * xp = opt->lbfgs.xp->data; // previous parameters
float * g = opt->lbfgs.g->data; // current gradient
@ -18785,6 +18785,12 @@ static enum ggml_opt_result ggml_opt_lbfgs(
float * lm_s = opt->lbfgs.lms->data;
float * lm_y = opt->lbfgs.lmy->data;
if (callback) {
// LBFG-S does not support learning rate -> ignore learning schedule
float sched = 0;
callback(callback_data, &sched);
}
// evaluate the function value and its gradient
{
ggml_opt_set_params(np, ps, x);
@ -18792,11 +18798,14 @@ static enum ggml_opt_result ggml_opt_lbfgs(
ggml_graph_reset (gf);
ggml_set_f32 (f->grad, 1.0f);
ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
ggml_graph_compute(gb, &cplan);
ggml_opt_get_grad(np, ps, g);
fx = ggml_get_f32_1d(f, 0);
opt->loss_before = fx;
opt->loss_after = fx;
}
// search direction = -gradient
@ -18851,7 +18860,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
ggml_vec_cpy_f32(nx, xp, x);
ggml_vec_cpy_f32(nx, gp, g);
ls = linesearch_backtracking(ctx, &params, nx, x, &fx, g, d, step, xp, f, gf, gb, np, ps);
ls = linesearch_backtracking(&params, nx, x, &fx, g, d, step, xp, f, gf, gb, &cplan, np, ps, callback, callback_data);
if (ls < 0) {
// linesearch failed - go back to the previous point and return
@ -18861,6 +18870,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
return ls;
}
opt->loss_after = fx;
ggml_vec_norm_f32(nx, &xnorm, x);
ggml_vec_norm_f32(nx, &gnorm, g);
@ -18918,7 +18929,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
// ys = y^t \cdot s -> 1 / \rho.
// yy = y^t \cdot y.
//
ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0] *nx]);
ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0]*nx]);
ggml_vec_dot_f32(nx, &yy, &lm_y[end[0]*nx], &lm_y[end[0]*nx]);
lm_ys[end[0]] = ys;
@ -18981,13 +18992,15 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
.adam = {
.n_iter = 10000,
.sched = 1.000f,
.decay = 0.001f,
.decay = 0.0f,
.decay_min_ndim = 2,
.alpha = 0.001f,
.beta1 = 0.9f,
.beta2 = 0.999f,
.eps = 1e-8f,
.eps_f = 1e-5f,
.eps_g = 1e-3f,
.gclip = 0.0f,
},
};
} break;
@ -19037,23 +19050,13 @@ GGML_API void ggml_opt_init(
switch (opt->params.type) {
case GGML_OPT_ADAM:
{
opt->adam.x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.g1 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.g2 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.m = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.mh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.vh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
opt->adam.pf = params.past > 0
? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.past)
: NULL;
ggml_set_zero(opt->adam.x);
ggml_set_zero(opt->adam.g1);
ggml_set_zero(opt->adam.g2);
ggml_set_zero(opt->adam.m);
ggml_set_zero(opt->adam.v);
ggml_set_zero(opt->adam.mh);
ggml_set_zero(opt->adam.vh);
if (opt->adam.pf) {
ggml_set_zero(opt->adam.pf);
}
@ -19137,7 +19140,7 @@ enum ggml_opt_result ggml_opt_resume(
*gf = ggml_build_forward (f);
*gb = ggml_build_backward(ctx, gf, true);
return ggml_opt_resume_g(ctx, opt, f, gf, gb);
return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL);
}
enum ggml_opt_result ggml_opt_resume_g(
@ -19145,7 +19148,9 @@ enum ggml_opt_result ggml_opt_resume_g(
struct ggml_opt_context * opt,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb) {
struct ggml_cgraph * gb,
ggml_opt_callback callback,
void * callback_data) {
// build forward + backward compute graphs
enum ggml_opt_result result = GGML_OPT_OK;
@ -19153,11 +19158,11 @@ enum ggml_opt_result ggml_opt_resume_g(
switch (opt->params.type) {
case GGML_OPT_ADAM:
{
result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb);
result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
} break;
case GGML_OPT_LBFGS:
{
result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb);
result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
} break;
}
@ -19612,7 +19617,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
// read the kv pairs
{
ctx->kv = GGML_ALIGNED_MALLOC(ctx->header.n_kv * sizeof(struct gguf_kv));
ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv));
for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
struct gguf_kv * kv = &ctx->kv[i];
@ -19695,7 +19700,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
// read the tensor infos
{
ctx->infos = GGML_ALIGNED_MALLOC(ctx->header.n_tensors * sizeof(struct gguf_tensor_info));
ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info));
for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
struct gguf_tensor_info * info = &ctx->infos[i];
@ -19896,7 +19901,7 @@ void gguf_free(struct gguf_context * ctx) {
}
}
GGML_ALIGNED_FREE(ctx->kv);
free(ctx->kv);
}
if (ctx->infos) {
@ -19908,7 +19913,7 @@ void gguf_free(struct gguf_context * ctx) {
}
}
GGML_ALIGNED_FREE(ctx->infos);
free(ctx->infos);
}
GGML_ALIGNED_FREE(ctx);

29
ggml.h
View file

@ -952,11 +952,11 @@ extern "C" {
// a - x
// b - dy
// TODO: update with configurable eps
GGML_API struct ggml_tensor * ggml_rms_norm_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * b,
float eps);
// A: n columns, m rows
// B: n columns, p rows (i.e. we transpose it internally)
@ -1612,7 +1612,8 @@ extern "C" {
struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
@ -1677,6 +1678,8 @@ extern "C" {
GGML_LINESEARCH_INVALID_PARAMETERS,
};
typedef void (*ggml_opt_callback)(void * data, float * sched);
// optimization parameters
//
// see ggml.c (ggml_opt_default_params) for default values
@ -1712,12 +1715,14 @@ extern "C" {
float sched; // schedule multiplier (fixed, decay or warmup)
float decay; // weight decay for AdamW, use 0.0f to disable
int decay_min_ndim; // minimum number of tensor dimension to apply weight decay
float alpha; // learning rate
float beta1;
float beta2;
float eps; // epsilon for numerical stability
float eps_f; // epsilon for convergence test
float eps_g; // epsilon for convergence test
float gclip; // gradient clipping
} adam;
// LBFGS parameters
@ -1745,14 +1750,12 @@ extern "C" {
bool just_initialized;
float loss_before;
float loss_after;
struct {
struct ggml_tensor * x; // view of the parameters
struct ggml_tensor * g1; // gradient
struct ggml_tensor * g2; // gradient squared
struct ggml_tensor * m; // first moment
struct ggml_tensor * v; // second moment
struct ggml_tensor * mh; // first moment hat
struct ggml_tensor * vh; // second moment hat
struct ggml_tensor * pf; // past function values
float fx_best;
float fx_prev;
@ -1789,10 +1792,10 @@ extern "C" {
// initialize optimizer context
GGML_API void ggml_opt_init(
struct ggml_context * ctx,
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_opt_params params,
int64_t nx);
struct ggml_opt_params params,
int64_t nx);
// continue optimizing the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt_resume(
@ -1806,7 +1809,9 @@ extern "C" {
struct ggml_opt_context * opt,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb);
struct ggml_cgraph * gb,
ggml_opt_callback callback,
void * callback_data);
//
// quantization

9
llama.cpp
View file

@ -6248,7 +6248,6 @@ const char * llama_print_system_info(void) {
}
void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) {
fprintf(stream, "\n");
fprintf(stream, "###########\n");
fprintf(stream, "# Timings #\n");
@ -6264,10 +6263,10 @@ void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) {
fprintf(stream, "n_eval: %d # number of tokens generated (excluding the first one)\n", ctx->n_eval);
fprintf(stream, "n_p_eval: %d # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
fprintf(stream, "n_sample: %d # number of sampled tokens\n", ctx->n_sample);
fprintf(stream, "t_eval_us: %ld # total microseconds spent generating tokens\n", ctx->t_eval_us);
fprintf(stream, "t_load_us: %ld # total microseconds spent loading the model\n", ctx->t_load_us);
fprintf(stream, "t_p_eval_us: %ld # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
fprintf(stream, "t_sample_us: %ld # total microseconds spent sampling\n", ctx->t_sample_us);
fprintf(stream, "t_eval_us: %" PRId64 " # total microseconds spent generating tokens\n", ctx->t_eval_us);
fprintf(stream, "t_load_us: %" PRId64 " # total microseconds spent loading the model\n", ctx->t_load_us);
fprintf(stream, "t_p_eval_us: %" PRId64 " # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
fprintf(stream, "t_sample_us: %" PRId64 " # total microseconds spent sampling\n", ctx->t_sample_us);
fprintf(stream, "ts_eval: %.2f # tokens / second during generation\n",
1.0e6 * ctx->n_eval / ctx->t_eval_us);
fprintf(stream, "ts_p_eval: %.2f # tokens / second during prompt processing\n",

52
tests/test-grad0.cpp
View file

@ -275,14 +275,14 @@ static bool check_gradient(
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
const float f0 = ggml_get_f32_1d(f, 0);
const double f0 = ggml_get_f32_1d(f, 0);
ggml_set_f32_1d(x[i], k, xm);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
const float f1 = ggml_get_f32_1d(f, 0);
const float g0 = (f0 - f1)/(2.0f*eps);
const double f1 = ggml_get_f32_1d(f, 0);
const double g0 = (f0 - f1)/(2.0*(double) eps);
ggml_set_f32_1d(x[i], k, x0);
@ -292,10 +292,10 @@ static bool check_gradient(
ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
const float g1 = ggml_get_f32_1d(x[i]->grad, k);
const double g1 = ggml_get_f32_1d(x[i]->grad, k);
const float error_abs = fabsf(g0 - g1);
const float error_rel = g0 != 0 ? fabsf(g0 - g1)/fabsf(g0) : 0;
const double error_abs = fabs(g0 - g1);
const double error_rel = g0 != 0 ? fabs(g0 - g1)/fabs(g0) : 0;
if (error_abs > max_error_abs || error_rel > max_error_rel) {
printf("%s: ndims=%d, i=%d, k=%d, x0=%f, xm=%f, xp=%f, f0=%f, f1=%f, g0=%f, g1=%f, eps=%f, error_abs=%f, error_rel=%f\n",
@ -531,7 +531,7 @@ int main(int argc, const char ** argv) {
struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqrt(ctx0, x[0]));
check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f);
check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f);
}
}
@ -1345,9 +1345,18 @@ int main(int argc, const char ** argv) {
x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
ggml_set_param(ctx0, x[0]);
struct ggml_tensor * f = ggml_sum(ctx0, ggml_soft_max(ctx0, x[0]));
float eps = 1e-6f;
// dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
// instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
struct ggml_tensor * f = ggml_sum(ctx0,
ggml_log(ctx0,
ggml_add1(ctx0,
ggml_scale(ctx0,
ggml_soft_max(ctx0, x[0]),
ggml_new_f32(ctx0, 1.0f - eps)),
ggml_new_f32(ctx0, eps))));
check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);
}
}
@ -1358,15 +1367,26 @@ int main(int argc, const char ** argv) {
int64_t ne2[4];
get_random_dims(ne2, 4);
for (int ndims = 1; ndims <= 3; ++ndims) {
x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
for (int ndims = 1; ndims <= 4; ++ndims) {
x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -0.1f, 0.1f);
x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
// the second argument to cross_entropy_loss must sum up to 1 for each row
int nr = ggml_nrows(x[1]);
int nc = ggml_nelements(x[1]) / nr;
for (int ir = 0; ir < nr; ++ir) {
float sum = 0;
for (int ic = 0; ic < nc; ++ic) {
sum += ((float *) x[1]->data)[ic + ir*nc];
}
for (int ic = 0; ic < nc; ++ic) {
((float *) x[1]->data)[ic + ir*nc] /= sum;
}
}
ggml_set_param(ctx0, x[0]);
struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
struct ggml_tensor * f = ggml_cross_entropy_loss(ctx0, x[0], x[1]);
check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
// finite differences regularly fails!
check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-4f, 1e-3f, INFINITY);
}
}
@ -1473,7 +1493,7 @@ int main(int argc, const char ** argv) {
struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
}
}
}
@ -1514,7 +1534,7 @@ int main(int argc, const char ** argv) {
struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
}
}
}