97 lines
3.3 KiB
C++
97 lines
3.3 KiB
C++
#include "arg.h"
|
|
#include "common.h"
|
|
#include "log.h"
|
|
#include "llama.h"
|
|
|
|
#include <cmath>
|
|
#include <cstdio>
|
|
#include <cstring>
|
|
#include <ctime>
|
|
#include <vector>
|
|
|
|
#if defined(_MSC_VER)
|
|
#pragma warning(disable: 4244 4267) // possible loss of data
|
|
#endif
|
|
|
|
int main(int argc, char ** argv) {
|
|
common_params params;
|
|
params.escape = false;
|
|
|
|
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_FINETUNE)) {
|
|
return 1;
|
|
}
|
|
|
|
if (params.use_mmap) {
|
|
LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n",
|
|
__func__);
|
|
params.use_mmap = false;
|
|
}
|
|
if (params.cache_type_k != GGML_TYPE_F32) {
|
|
LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
|
|
params.cache_type_k = GGML_TYPE_F32;
|
|
}
|
|
if (params.cache_type_v != GGML_TYPE_F32) {
|
|
LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
|
|
params.cache_type_v = GGML_TYPE_F32;
|
|
}
|
|
|
|
common_init();
|
|
llama_backend_init();
|
|
llama_numa_init(params.numa);
|
|
// load the model and apply lora adapter, if any
|
|
common_init_result llama_init = common_init_from_params(params);
|
|
llama_model_ptr & model = llama_init.model;
|
|
llama_context_ptr & ctx = llama_init.context;
|
|
|
|
if (model == NULL) {
|
|
LOG_ERR("%s: unable to load model\n", __func__);
|
|
return 1;
|
|
}
|
|
|
|
// print system information
|
|
{
|
|
LOG_INF("\n");
|
|
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
|
|
}
|
|
|
|
std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
|
|
ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get()) / 2);
|
|
|
|
struct lr_opt & lr = params.lr;
|
|
LOG_INF("-optimizer %s -lr0 %.2g -wd %.2g -lr-min %.2g -min-epochs %.2g -epochs %d -period %.2g -val %.2g\n",
|
|
ggml_opt_optimizer_name(params.optimizer), (double) lr.lr0, (double) lr.wd, (double) lr.lr_min, (double) lr.decay_epochs,
|
|
(unsigned) lr.epochs, (double) params.n_batch / params.n_ubatch, (double) params.val_split);
|
|
|
|
struct llama_opt_params lopt_params{
|
|
/*n_ctx_train =*/0,
|
|
/*param_filter =*/llama_opt_param_filter_all,
|
|
/*param_filter_ud =*/nullptr,
|
|
/*get_opt_pars =*/common_opt_lr_pars,
|
|
/*get_opt_pars_ud =*/¶ms.lr,
|
|
/*optimizer_type =*/params.optimizer,
|
|
};
|
|
llama_opt_init(ctx.get(), model.get(), lopt_params);
|
|
|
|
const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - params.val_split);
|
|
|
|
ggml_opt_result_t result_train = ggml_opt_result_init();
|
|
ggml_opt_result_t result_eval = ggml_opt_result_init();
|
|
|
|
for (lr.epoch = 0; lr.epoch < lr.epochs; ++lr.epoch) {
|
|
llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
|
|
ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
|
|
fprintf(stderr, "\n");
|
|
|
|
ggml_opt_result_reset(result_train);
|
|
ggml_opt_result_reset(result_eval);
|
|
}
|
|
ggml_opt_result_free(result_train);
|
|
ggml_opt_result_free(result_eval);
|
|
|
|
llama_model_save_to_file(model.get(), params.out_file.c_str());
|
|
|
|
llama_backend_free();
|
|
|
|
return 0;
|
|
}
|