CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

finetune: SGD optimizer, more CLI args (#13873)

Browse Source 
* examples/finetune -opt SGD (stochastic gradient descent) memory opt

add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.

support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)

llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)

(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val:   [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00

SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val:   [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)

note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')

-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.

note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence

new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)

cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)

since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)

test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values);  tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)

* Vulkan: Implement GGML_OP_OPT_STEP_SGD

* tests: Fix OPT_STEP_SGD test-backend-ops

* SGD op param store weight-decay and not 1-alpha*wd

* minor + cosmetic changes

* fix vulkan sgd

* try CI fix

---------

Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Jonathan Graehl
2025-08-14 03:03:57 -07:00
committed by GitHub
parent 3ea913f1ce
commit 5cdb27e091
24 changed files with 718 additions and 187 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
common/arg.cpp
View File

@@ -1238,6 +1238,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
common_params_print_completion(ctx_arg);
exit(0);
}
params.lr.init();
} catch (const std::invalid_argument & ex) {
fprintf(stderr, "%s\n", ex.what());
ctx_arg.params = params_org;
@@ -2688,7 +2689,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.out_file = value;
}
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
add_opt(common_arg(
{"-ofreq", "--output-frequency"}, "N",
string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -3566,5 +3567,51 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(
common_arg({ "-lr", "--learning-rate" }, "ALPHA",
string_format(
"adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
(double) params.lr.lr0),
[](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(
common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
string_format(
"(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
(double) params.lr.lr_min),
[](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(
common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
string_format(
"(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
(double) params.lr.decay_epochs),
[](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg(
{ "-wd", "--weight-decay" }, "WD",
string_format(
"adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
(double) params.lr.wd),
[](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
string_format("fraction of data to use as validation set for training (default: %.2g).",
(double) params.val_split),
[](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-epochs", "--epochs" }, "N",
string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
[](common_params & params, int epochs) { params.lr.epochs = epochs; })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
[](common_params & params, const std::string & name) {
params.optimizer = common_opt_get_optimizer(name.c_str());
if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
}
})
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
return ctx_arg;
}