1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
|
#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_PERPLEXITY)) {
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());
}
constexpr float val_split = 0.05f;
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 ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
optimizer_params.adamw.alpha = 1e-7f; // learning rate
struct llama_opt_params lopt_params {
/*n_ctx_train =*/ 0,
/*param_filter =*/ llama_opt_param_filter_all,
/*param_filter_ud =*/ nullptr,
/*get_opt_pars =*/ ggml_opt_get_constant_optimizer_params,
/*get_opt_pars_ud =*/ &optimizer_params,
};
llama_opt_init(ctx.get(), model.get(), lopt_params);
const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
ggml_opt_result_t result_train = ggml_opt_result_init();
ggml_opt_result_t result_eval = ggml_opt_result_init();
for (int epoch = 0; epoch < 2; ++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(), "finetuned-model.gguf");
llama_backend_free();
return 0;
}
|
