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

gguf : deduplicate (#2629)

Browse source 
* gguf : better type names

* dedup : CPU + Metal is working

* ggml : fix warnings about unused results

* llama.cpp : fix line feed and compiler warning

* llama : fix strncpy warning + note token_to_str does not write null

* llama : restore the original load/save session implementation

Will migrate this to GGUF in the future

* convert-llama-h5-to-gguf.py : support alt ctx param name

* ggml : assert when using ggml_mul with non-F32 src1

* examples : dedup simple

---------

Co-authored-by: klosax <131523366+klosax@users.noreply.github.com>
This commit is contained in:
Georgi Gerganov 2023-08-16 19:25:29 +03:00 • committed by GitHub
parent 758ff1bbb5
commit 88b5769487
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
21 changed files with 1630 additions and 7398 deletions
Download patch file Download diff file Expand all files Collapse all files

16
examples/common.cpp
View file

@ -170,18 +170,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
break;
}
params.n_ctx = std::stoi(argv[i]);
} else if (arg == "-gqa" || arg == "--gqa") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.n_gqa = std::stoi(argv[i]);
} else if (arg == "-eps" || arg == "--rms-norm-eps") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rms_norm_eps = std::stof(argv[i]);
} else if (arg == "--rope-freq-base") {
if (++i >= argc) {
invalid_param = true;
@ -546,8 +534,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stdout, " -gqa N, --gqa N grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
fprintf(stdout, " -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
fprintf(stdout, " --top-k N top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
fprintf(stdout, " --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
fprintf(stdout, " --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
@ -638,8 +624,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
lparams.n_ctx = params.n_ctx;
lparams.n_batch = params.n_batch;
lparams.n_gqa = params.n_gqa;
lparams.rms_norm_eps = params.rms_norm_eps;
lparams.n_gpu_layers = params.n_gpu_layers;
lparams.main_gpu = params.main_gpu;
lparams.tensor_split = params.tensor_split;

2
examples/common.h
View file

@ -23,14 +23,12 @@ struct gpt_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_gqa = 1; // grouped-query attention factor (TODO: move to hparams)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_gpu_layers = 0; // number of layers to store in VRAM
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; // rms norm epsilon
float rope_freq_base = 10000.0f; // RoPE base frequency
float rope_freq_scale = 1.0f; // RoPE frequency scaling factor

144
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -1,5 +1,6 @@
#include "ggml.h"
#include "llama.h"
#include <unordered_map>
#include <vector>
#include <cassert>
@ -502,7 +503,7 @@ bool is_ggml_file(const char *filename) {
return false;
}
uint32_t magic = file.read_u32();
return magic == LLAMA_FILE_MAGIC;
return magic == GGUF_MAGIC;
}
void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
@ -590,75 +591,80 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
if (file.fp == NULL) {
return;
}
// write_magic
file.write_u32(LLAMA_FILE_MAGIC); // magic
file.write_u32(LLAMA_FILE_VERSION); // version
// write_hparams
file.write_u32(model->hparams.n_vocab);
file.write_u32(model->hparams.n_embd);
file.write_u32(model->hparams.n_mult);
file.write_u32(model->hparams.n_head);
file.write_u32(model->hparams.n_layer);
file.write_u32(model->hparams.n_rot);
file.write_u32(LLAMA_FTYPE_ALL_F32);
// write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
uint32_t n_vocab = model->hparams.n_vocab;
for (uint32_t i = 0; i < n_vocab; i++) {
const auto & token_score = vocab->id_to_token.at(i);
file.write_u32((uint32_t) token_score.tok.size());
file.write_raw(token_score.tok.data(), token_score.tok.size());
file.write_raw(&token_score.score, sizeof(token_score.score));
}
// stuff AK weights into GG weights one by one.
// w->token_embedding_table -> model->tok_embeddings
// float* -> struct ggml_tensor
stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
//print_row(model->norm, 0);
// for rms-att-weight
int row_length = model->hparams.n_embd;
const auto & hparams = model->hparams;
//int n_ff = model->hparams.n_embd;
int n_ff = get_n_ff(&hparams);
for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
auto & layer = model->layers[i];
// 1d
stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]);
// from 3d matrix layer x dim x dim to 2d matrix dim x dim
stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]);
stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]);
stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]);
}
// write tensors
write_tensor(&file, model->tok_embeddings);
write_tensor(&file, model->norm);
write_tensor(&file, model->output); // ?
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
auto & layer = model->layers[i];
write_tensor(&file, layer.attention_norm);
write_tensor(&file, layer.wq);
write_tensor(&file, layer.wk);
write_tensor(&file, layer.wv);
write_tensor(&file, layer.wo);
write_tensor(&file, layer.ffn_norm);
write_tensor(&file, layer.w1);
write_tensor(&file, layer.w2);
write_tensor(&file, layer.w3);
}
#pragma message("TODO: implement file saving using gguf")
(void) vocab;
(void) model;
(void) w;
// // write_magic
// file.write_u32(LLAMA_FILE_MAGIC); // magic
// file.write_u32(LLAMA_FILE_VERSION); // version
// // write_hparams
// file.write_u32(model->hparams.n_vocab);
// file.write_u32(model->hparams.n_embd);
// file.write_u32(model->hparams.n_mult);
// file.write_u32(model->hparams.n_head);
// file.write_u32(model->hparams.n_layer);
// file.write_u32(model->hparams.n_rot);
// file.write_u32(LLAMA_FTYPE_ALL_F32);
//
// // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
// uint32_t n_vocab = model->hparams.n_vocab;
// for (uint32_t i = 0; i < n_vocab; i++) {
// const auto & token_score = vocab->id_to_token.at(i);
// file.write_u32((uint32_t) token_score.tok.size());
// file.write_raw(token_score.tok.data(), token_score.tok.size());
// file.write_raw(&token_score.score, sizeof(token_score.score));
// }
//
// // stuff AK weights into GG weights one by one.
// // w->token_embedding_table -> model->tok_embeddings
// // float* -> struct ggml_tensor
// stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
// stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
//
// stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
// //print_row(model->norm, 0);
//
// // for rms-att-weight
// int row_length = model->hparams.n_embd;
// const auto & hparams = model->hparams;
// //int n_ff = model->hparams.n_embd;
// int n_ff = get_n_ff(&hparams);
//
// for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
// auto & layer = model->layers[i];
// // 1d
// stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
// stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]);
//
// // from 3d matrix layer x dim x dim to 2d matrix dim x dim
// stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]);
// stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]);
// stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]);
// stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]);
//
// stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]);
// stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]);
// stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]);
// }
// // write tensors
// write_tensor(&file, model->tok_embeddings);
// write_tensor(&file, model->norm);
// write_tensor(&file, model->output); // ?
// for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
// auto & layer = model->layers[i];
//
// write_tensor(&file, layer.attention_norm);
// write_tensor(&file, layer.wq);
// write_tensor(&file, layer.wk);
// write_tensor(&file, layer.wv);
// write_tensor(&file, layer.wo);
// write_tensor(&file, layer.ffn_norm);
// write_tensor(&file, layer.w1);
// write_tensor(&file, layer.w2);
// write_tensor(&file, layer.w3);
// }
}
struct train_params get_default_train_params() {

129
examples/gguf/gguf-llama-simple.cpp
View file

@ -1,129 +0,0 @@
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include "common.h"
#include "gguf-llama.h"
#include "build-info.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
int main(int argc, char ** argv) {
gpt_params params;
if (argc == 1 || argv[1][0] == '-') {
printf("usage: %s MODEL_PATH [PROMPT]\n" , argv[0]);
return 1 ;
}
if (argc >= 2) {
params.model = argv[1];
}
if (argc >= 3) {
params.prompt = argv[2];
}
if (params.prompt.empty()) {
params.prompt = "Hello my name is";
}
// init LLM
llama_backend_init(params.numa);
llama_context_params ctx_params = llama_context_default_params();
llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true);
const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
if ((int) tokens_list.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_str(ctx, id).c_str());
}
fflush(stderr);
// main loop
// The LLM keeps a contextual cache memory of previous token evaluation.
// Usually, once this cache is full, it is required to recompute a compressed context based on previous
// tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
// example, we will just stop the loop once this cache is full or once an end of stream is detected.
const int n_gen = std::min(32, max_context_size);
while (llama_get_kv_cache_token_count(ctx) < n_gen) {
// evaluate the transformer
if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return 1;
}
tokens_list.clear();
// sample the next token
llama_token new_token_id = 0;
auto logits = llama_get_logits(ctx);
auto n_vocab = llama_n_vocab(ctx);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
// is it an end of stream ?
if (new_token_id == llama_token_eos()) {
fprintf(stderr, " [end of text]\n");
break;
}
// print the new token :
printf("%s", llama_token_to_str(ctx, new_token_id).c_str());
fflush(stdout);
// push this new token for next evaluation
tokens_list.push_back(new_token_id);
}
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
return 0;
}

2
examples/gguf/gguf.cpp
View file

@ -1,5 +1,5 @@
#include "ggml.h"
#include "gguf-llama.h"
#include "llama.h"
#include <cstdio>
#include <cinttypes>

6
examples/main/main.cpp
View file

@ -266,9 +266,6 @@ int main(int argc, char ** argv) {
params.interactive = true;
}
// determine newline token
auto llama_token_newline = ::llama_tokenize(ctx, "\n", false);
if (params.verbose_prompt) {
fprintf(stderr, "\n");
fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
@ -778,8 +775,7 @@ int main(int argc, char ** argv) {
if (grammar != NULL) {
llama_grammar_free(grammar);
std::vector<const llama_grammar_element *> grammar_rules(
parsed_grammar.c_rules());
std::vector<const llama_grammar_element *> grammar_rules( parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(),
parsed_grammar.symbol_ids.at("root"));

8
examples/quantize/quantize.cpp
View file

@ -68,10 +68,10 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
}
// usage:
// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
//
void usage(const char * executable) {
fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.bin [model-quant.bin] type [nthreads]\n\n", executable);
fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
fprintf(stderr, " --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
fprintf(stderr, " --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
fprintf(stderr, "\nAllowed quantization types:\n");
@ -118,8 +118,8 @@ int main(int argc, char ** argv) {
if (pos != std::string::npos) {
fpath = fname_inp.substr(0, pos + 1);
}
// export as [inp path]/ggml-model-[ftype].bin
fname_out = fpath + "ggml-model-" + ftype_str + ".bin";
// export as [inp path]/ggml-model-[ftype].gguf
fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
arg_idx++;
}
else {

1
examples/save-load-state/save-load-state.cpp
View file

@ -26,7 +26,6 @@ int main(int argc, char ** argv) {
auto lparams = llama_context_default_params();
lparams.n_ctx = params.n_ctx;
lparams.n_gqa = params.n_gqa;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
lparams.use_mmap = params.use_mmap;

19
examples/server/server.cpp
View file

@ -651,8 +651,6 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
fprintf(stdout, " -v, --verbose verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
fprintf(stdout, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " -gqa N, --gqa N grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
fprintf(stdout, " -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
fprintf(stdout, " --rope-freq-base N RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
fprintf(stdout, " --rope-freq-scale N RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
@ -773,23 +771,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.n_ctx = std::stoi(argv[i]);
}
else if (arg == "-gqa" || arg == "--gqa")
{
if (++i >= argc)
{
invalid_param = true;
break;
}
params.n_gqa = std::stoi(argv[i]);
}
else if (arg == "-eps" || arg == "--rms-norm-eps") {
if (++i >= argc)
{
invalid_param = true;
break;
}
params.rms_norm_eps = std::stof(argv[i]);
}
else if (arg == "--rope-freq-base")
{
if (++i >= argc)

18
examples/simple/simple.cpp
View file

@ -36,16 +36,17 @@ int main(int argc, char ** argv) {
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
llama_context_params ctx_params = llama_context_default_params();
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
// tokenize the prompt
std::vector<llama_token> tokens_list;
@ -54,7 +55,7 @@ int main(int argc, char ** argv) {
const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
if ((int)tokens_list.size() > max_tokens_list_size) {
if ((int) tokens_list.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
return 1;
}
@ -74,7 +75,9 @@ int main(int argc, char ** argv) {
// tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
// example, we will just stop the loop once this cache is full or once an end of stream is detected.
while (llama_get_kv_cache_token_count( ctx ) < max_context_size) {
const int n_gen = std::min(32, max_context_size);
while (llama_get_kv_cache_token_count(ctx) < n_gen) {
// evaluate the transformer
if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
@ -114,7 +117,6 @@ int main(int argc, char ** argv) {
// push this new token for next evaluation
tokens_list.push_back(new_token_id);
}
llama_free(ctx);
@ -122,5 +124,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
return 0;
}

77
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -17,7 +17,7 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
static const float rms_norm_eps = 1e-5f;
struct random_normal_distribution {
std::mt19937 gen;
@ -2612,42 +2612,45 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
return;
}
// write_magic
file.write_u32(LLAMA_FILE_MAGIC); // magic
file.write_u32(LLAMA_FILE_VERSION); // version
// write_hparams
file.write_u32(model->hparams.n_vocab);
file.write_u32(model->hparams.n_embd);
file.write_u32(model->hparams.n_mult);
file.write_u32(model->hparams.n_head);
file.write_u32(model->hparams.n_layer);
file.write_u32(model->hparams.n_rot);
file.write_u32(LLAMA_FTYPE_ALL_F32);
// write_vocab
uint32_t n_vocab = model->hparams.n_vocab;
for (uint32_t i = 0; i < n_vocab; i++) {
const auto & token_score = vocab->id_to_token.at(i);
file.write_u32((uint32_t) token_score.tok.size());
file.write_raw(token_score.tok.data(), token_score.tok.size());
file.write_raw(&token_score.score, sizeof(token_score.score));
}
// write tensors
write_tensor(&file, model->tok_embeddings);
write_tensor(&file, model->norm);
write_tensor(&file, model->output);
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
auto & layer = model->layers[i];
write_tensor(&file, layer.attention_norm);
write_tensor(&file, layer.wq);
write_tensor(&file, layer.wk);
write_tensor(&file, layer.wv);
write_tensor(&file, layer.wo);
write_tensor(&file, layer.ffn_norm);
write_tensor(&file, layer.w1);
write_tensor(&file, layer.w2);
write_tensor(&file, layer.w3);
}
#pragma message("TODO: implement file saving using gguf")
(void) vocab;
(void) model;
// // write_magic
// file.write_u32(LLAMA_FILE_MAGIC); // magic
// file.write_u32(LLAMA_FILE_VERSION); // version
// // write_hparams
// file.write_u32(model->hparams.n_vocab);
// file.write_u32(model->hparams.n_embd);
// file.write_u32(model->hparams.n_mult);
// file.write_u32(model->hparams.n_head);
// file.write_u32(model->hparams.n_layer);
// file.write_u32(model->hparams.n_rot);
// file.write_u32(LLAMA_FTYPE_ALL_F32);
// // write_vocab
// uint32_t n_vocab = model->hparams.n_vocab;
// for (uint32_t i = 0; i < n_vocab; i++) {
// const auto & token_score = vocab->id_to_token.at(i);
// file.write_u32((uint32_t) token_score.tok.size());
// file.write_raw(token_score.tok.data(), token_score.tok.size());
// file.write_raw(&token_score.score, sizeof(token_score.score));
// }
// // write tensors
// write_tensor(&file, model->tok_embeddings);
// write_tensor(&file, model->norm);
// write_tensor(&file, model->output);
// for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
// auto & layer = model->layers[i];
//
// write_tensor(&file, layer.attention_norm);
// write_tensor(&file, layer.wq);
// write_tensor(&file, layer.wk);
// write_tensor(&file, layer.wv);
// write_tensor(&file, layer.wo);
// write_tensor(&file, layer.ffn_norm);
// write_tensor(&file, layer.w1);
// write_tensor(&file, layer.w2);
// write_tensor(&file, layer.w3);
// }
}
float cosine_decay(const int decay_steps, const float alpha, int step) {