summaryrefslogtreecommitdiff
path: root/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp')
-rw-r--r--examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp180
1 files changed, 91 insertions, 89 deletions
diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
index 1a238c4d..469d6e3d 100644
--- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
@@ -1,5 +1,6 @@
#include "ggml.h"
#include "llama.h"
+
#include <unordered_map>
#include <vector>
#include <cassert>
@@ -138,14 +139,16 @@ void print_sample_weights(TransformerWeights *w){
struct llama_vocab {
using id = int32_t;
using token = std::string;
+ using ttype = llama_token_type;
- struct token_score {
- token tok;
+ struct token_data {
+ token text;
float score;
+ ttype type;
};
std::unordered_map<token, id> token_to_id;
- std::vector<token_score> id_to_token;
+ std::vector<token_data> id_to_token;
};
struct my_llama_hparams {
@@ -502,7 +505,7 @@ 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) {
@@ -515,36 +518,30 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
- std::vector<const char *> strings;
- std::vector<float> scores;
- int n_vocab = llama_n_vocab(lctx);
- strings.resize(n_vocab, NULL);
- scores.resize(n_vocab, 0);
- n_vocab = llama_get_vocab(lctx, strings.data(), scores.data(), n_vocab);
- GGML_ASSERT(n_vocab == llama_n_vocab(lctx));
+ const int n_vocab = llama_n_vocab(lctx);
vocab->id_to_token.resize(n_vocab);
for (int i=0; i<n_vocab; ++i) {
- std::string tok = std::string(strings[i]);
- float score = scores[i];
- vocab->id_to_token[i].tok = tok;
- vocab->id_to_token[i].score = score;
- vocab->token_to_id.emplace(tok, i);
+ vocab->id_to_token[i].text = llama_token_get_text(lctx, i);
+ vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
+ vocab->id_to_token[i].type = llama_token_get_type(lctx, i);
+ vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
}
llama_free(lctx);
llama_free_model(lmodel);
} else { // assume llama2.c vocabulary
printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
llama_file file(filename, "rb");
- uint32_t n_vocab = config->vocab_size;
+ const int n_vocab = config->vocab_size;
/* uint32_t max_token_length = */ file.read_u32(); // unused
vocab->id_to_token.resize(n_vocab);
- for (uint32_t i=0; i<n_vocab; ++i) {
+ for (int i=0; i<n_vocab; ++i) {
float_t score = file.read_f32();
uint32_t len = file.read_u32();
- std::string tok = file.read_string(len);
- vocab->id_to_token[i].tok = tok;
+ std::string text = file.read_string(len);
+ vocab->id_to_token[i].text = text;
vocab->id_to_token[i].score = score;
- vocab->token_to_id.emplace(tok, i);
+ vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
+ vocab->token_to_id.emplace(text, i);
}
}
}
@@ -590,75 +587,80 @@ 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_data = vocab->id_to_token.at(i);
+// file.write_u32((uint32_t) token_data.tok.size());
+// file.write_raw(token_data.tok.data(), token_data.tok.size());
+// file.write_raw(&token_data.score, sizeof(token_data.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() {
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-15 14:27:28 +0000