CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +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;
}

54
common/common.cpp
View File

@@ -41,6 +41,7 @@
#endif
#include <locale>
#include <windows.h>
#include <string.h>
#include <fcntl.h>
#include <io.h>
#else
@@ -1565,3 +1566,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
return result;
}
ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
const lr_opt & d = *(lr_opt *) userdata;
result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
result.sgd.wd = result.adamw.wd = d.wd;
return result;
}
// TODO make all command line args case-insensitive
static inline bool eq_case_insensitive(char const* a, char const* b) {
return !
#if defined(_MSC_VER)
_stricmp
#else
strcasecmp
#endif // defined(_MSC_VER)
(a, b);
}
enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
if (eq_case_insensitive("adamw", n)) {
return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
}
if (eq_case_insensitive("sgd", n)) {
return GGML_OPT_OPTIMIZER_TYPE_SGD;
}
return GGML_OPT_OPTIMIZER_TYPE_COUNT;
}
// TODO simplify to use just log and exp
static float const k_log_2 = std::log(2.f);
void lr_opt::init() {
if (lr_min > 0 && lr_min < lr0) {
float nhalf = std::log(lr0 / lr_min) / k_log_2;
float e = epochs;
if (decay_epochs > 0 && decay_epochs < e) {
e = decay_epochs;
} else {
decay_epochs = e;
}
scale_epoch = nhalf / e;
}
}
float lr_opt::get_lr(float epoch) const {
float r = lr_min <= 0 ? lr0 :
epoch >= decay_epochs ? lr_min :
lr0 * std::pow(0.5f, epoch * scale_epoch);
LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
return r;
}

35
common/common.h
View File

@@ -2,14 +2,17 @@
#pragma once
#include "llama-cpp.h"
#include <set>
#include <sstream>
#include <string>
#include <string_view>
#include <vector>
#include <map>
#include <sstream>
#include <cmath>
#include "ggml-opt.h"
#include "llama-cpp.h"
#ifdef _WIN32
#define DIRECTORY_SEPARATOR '\\'
@@ -82,6 +85,7 @@ enum llama_example {
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_TTS,
LLAMA_EXAMPLE_DIFFUSION,
LLAMA_EXAMPLE_FINETUNE,
LLAMA_EXAMPLE_COUNT,
};
@@ -243,6 +247,25 @@ enum common_reasoning_format {
COMMON_REASONING_FORMAT_GRANITE, // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
};
struct lr_opt {
float lr0 = 1e-5; // learning rate at first epoch
float lr_min = -1;
float decay_epochs = -1; // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
float scale_epoch = 0;
float wd = 0;
unsigned epochs = 2;
unsigned epoch; // set by optimizer outer (epochs) loop
// learning rate decay - constant LR per epoch only for now
float get_lr(float e) const;
float get_lr() const { return get_lr(epoch); }
// must call after arg parse, before get_lr
void init();
};
struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 4096; // context size
@@ -377,6 +400,11 @@ struct common_params {
bool no_mmproj = false; // explicitly disable multimodal model
std::vector<std::string> image; // path to image file(s)
// finetune
struct lr_opt lr;
enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
float val_split = 0.05f; // fraction of the data used for the validation set
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@@ -704,3 +732,6 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
//
ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
// "adamw" or "sgd" (case insensitive)
enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);