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

llama2.c: direct gguf output (WIP)

Browse source
This commit is contained in:
ochafik 2023-08-23 21:08:00 +01:00
parent cf658adc83
commit 5132130af7
2 changed files with 118 additions and 111 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/convert-llama2c-to-ggml/README.md
View file

@ -19,11 +19,7 @@ options:
An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows: An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin` `$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually:
`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output stories42M.gguf.bin`
Now you can use the model with a command like: Now you can use the model with a command like:

223
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -13,6 +13,20 @@
#include <algorithm> #include <algorithm>
#include <string> #include <string>
// tensor names
#define TN_TOKEN_EMBD "token_embd.weight"
#define TN_OUTPUT_NORM "output_norm.weight"
#define TN_OUTPUT "output.weight"
#define TN_ATTN_NORM "blk.%d.attn_norm.weight"
#define TN_ATTN_Q "blk.%d.attn_q.weight"
#define TN_ATTN_K "blk.%d.attn_k.weight"
#define TN_ATTN_V "blk.%d.attn_v.weight"
#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
#define TN_FFN_NORM "blk.%d.ffn_norm.weight"
#define TN_FFN_GATE "blk.%d.ffn_gate.weight"
#define TN_FFN_DOWN "blk.%d.ffn_down.weight"
#define TN_FFN_UP "blk.%d.ffn_up.weight"
#if defined(_MSC_VER) #if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data #pragma warning(disable: 4244 4267) // possible loss of data
#endif #endif
@ -183,6 +197,7 @@ struct my_llama_hparams {
uint32_t n_vocab = 32000; uint32_t n_vocab = 32000;
uint32_t n_ctx = 512; // this is provided as user input? uint32_t n_ctx = 512; // this is provided as user input?
uint32_t n_embd = 4096; uint32_t n_embd = 4096;
uint32_t n_ff = 11008;
uint32_t n_mult = 4; uint32_t n_mult = 4;
uint32_t n_head = 32; uint32_t n_head = 32;
uint32_t n_layer = 32; uint32_t n_layer = 32;
@ -287,7 +302,7 @@ void print_params(struct my_llama_hparams * params) {
printf("%s: n_embd: %d\n", __func__, params->n_embd); printf("%s: n_embd: %d\n", __func__, params->n_embd);
printf("%s: n_mult: %d\n", __func__, params->n_mult); printf("%s: n_mult: %d\n", __func__, params->n_mult);
printf("%s: n_head: %d\n", __func__, params->n_head); printf("%s: n_head: %d\n", __func__, params->n_head);
printf("%s: n_ff: %d\n", __func__, get_n_ff(params)); printf("%s: n_ff: %d\n", __func__, params->n_ff);
printf("%s: n_layer: %d\n", __func__, params->n_layer); printf("%s: n_layer: %d\n", __func__, params->n_layer);
printf("%s: n_rot: %d\n", __func__, params->n_rot); printf("%s: n_rot: %d\n", __func__, params->n_rot);
} }
@ -299,7 +314,7 @@ void init_model(struct my_llama_model * model) {
const uint32_t n_layer = hparams.n_layer; const uint32_t n_layer = hparams.n_layer;
const uint32_t n_vocab = hparams.n_vocab; const uint32_t n_vocab = hparams.n_vocab;
const uint32_t n_ff = get_n_ff(&hparams); const uint32_t n_ff = hparams.n_ff;
struct ggml_context * ctx = model->ctx; struct ggml_context * ctx = model->ctx;
model->train_its = 0; model->train_its = 0;
@ -503,30 +518,6 @@ struct llama_file {
} }
}; };
void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
if (tensor == NULL) {
file->write_u32(0);
file->write_u32(0);
file->write_u32(GGML_TYPE_F32);
file->seek((0-file->tell()) & 31, SEEK_CUR);
return;
}
const char * name = ggml_get_name(tensor);
uint32_t name_len = strlen(name);
uint32_t nd = tensor->n_dims;
uint32_t ne[4] = { (uint32_t)tensor->ne[0],
(uint32_t)tensor->ne[1],
(uint32_t)tensor->ne[2],
(uint32_t)tensor->ne[3] };
file->write_u32(nd);
file->write_u32(name_len);
file->write_u32(tensor->type);
file->write_raw(ne, sizeof(ne[0]) * nd);
file->write_raw(name, name_len);
file->seek((0-file->tell()) & 31, SEEK_CUR);
file->write_raw(tensor->data, ggml_nbytes(tensor));
}
bool is_ggml_file(const char *filename) { bool is_ggml_file(const char *filename) {
llama_file file(filename, "rb"); llama_file file(filename, "rb");
if (file.size < 4) { if (file.size < 4) {
@ -537,48 +528,26 @@ bool is_ggml_file(const char *filename) {
} }
void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
#pragma message("TODO: implement reading vocabulary using gguf") // assume llama2.c vocabulary
// // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
// if (is_ggml_file(filename)) { llama_file file(filename, "rb");
// const int n_vocab = config->vocab_size;
// struct llama_context_params llama_params = llama_context_default_params(); /* uint32_t max_token_length = */ file.read_u32(); // unused
// llama_params.vocab_only = true; vocab->id_to_token.resize(n_vocab);
// for (int i=0; i<n_vocab; ++i) {
// struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params); float_t score = file.read_f32();
// struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params); uint32_t len = file.read_u32();
// std::string text = file.read_string(len);
// const int n_vocab = llama_n_vocab(lctx); // Special-case handling of <0xXX> single byte tokens.
// vocab->id_to_token.resize(n_vocab); char byte_val;
// for (int i=0; i<n_vocab; ++i) { if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
// vocab->id_to_token[i].text = llama_token_get_text(lctx, i); char cstr[2] = { byte_val, 0 };
// vocab->id_to_token[i].score = llama_token_get_score(lctx, i); text = cstr;
// 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");
const int n_vocab = config->vocab_size;
/* uint32_t max_token_length = */ file.read_u32(); // unused
vocab->id_to_token.resize(n_vocab);
for (int i=0; i<n_vocab; ++i) {
float_t score = file.read_f32();
uint32_t len = file.read_u32();
std::string text = file.read_string(len);
// Special-case handling of <0xXX> single byte tokens.
char byte_val;
if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
char cstr[2] = { byte_val, 0 };
text = cstr;
}
vocab->id_to_token[i].text = text;
vocab->id_to_token[i].score = score;
vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
vocab->token_to_id.emplace(text, i);
} }
vocab->id_to_token[i].text = text;
vocab->id_to_token[i].score = score;
vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
vocab->token_to_id.emplace(text, i);
} }
} }
@ -619,33 +588,6 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar
} }
void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) { void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
struct llama_file file(filename, "wb");
if (file.fp == NULL) {
return;
}
#pragma message("TODO: implement file saving using gguf")
// write_magic
file.write_u32(LLAMA_FILE_MAGIC_GGJT); // magic
file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // 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.text.size());
file.write_raw(token_data.text.data(), token_data.text.size());
file.write_raw(&token_data.score, sizeof(token_data.score));
}
// stuff AK weights into GG weights one by one. // stuff AK weights into GG weights one by one.
// w->token_embedding_table -> model->tok_embeddings // w->token_embedding_table -> model->tok_embeddings
// float* -> struct ggml_tensor // float* -> struct ggml_tensor
@ -677,23 +619,91 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]); 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]); stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]);
} }
struct gguf_context * ctx = gguf_init_empty();
std::vector<const char*> tokens;
std::transform(vocab->id_to_token.begin(), vocab->id_to_token.end(), std::back_inserter(tokens),
[](const llama_vocab::token_data & token_data) { return token_data.text.c_str(); });
gguf_set_arr_str(ctx, "tokenizer.ggml.tokens", tokens.data(), tokens.size());
std::vector<float> scores;
std::transform(vocab->id_to_token.begin(), vocab->id_to_token.end(), std::back_inserter(scores),
[](const llama_vocab::token_data & token_data) { return token_data.score; });
gguf_set_arr_data(ctx, "tokenizer.ggml.scores", GGUF_TYPE_FLOAT32, scores.data(), scores.size());
std::vector<int> token_types;
for (size_t i = 0; i < vocab->id_to_token.size(); ++i) {
// TODO: Refine this.
token_types.push_back(LLAMA_TOKEN_TYPE_UNDEFINED);
}
gguf_set_arr_data(ctx, "tokenizer.ggml.token_type", GGUF_TYPE_INT32, token_types.data(), token_types.size());
gguf_set_val_str(ctx, "tokenizer.ggml.model", "llama");
gguf_set_val_str(ctx, "general.name", "llama2.c");
gguf_set_val_str(ctx, "general.architecture", "llama");
// special tokens
// gguf_set_val_u32(ctx, "tokenizer.ggml.bos_token_id", ...);
// gguf_set_val_u32(ctx, "tokenizer.ggml.eos_token_id", ...);
// gguf_set_val_u32(ctx, "tokenizer.ggml.unknown_token_id", ...);
// gguf_set_val_u32(ctx, "tokenizer.ggml.separator_token_id", ...);
// gguf_set_val_u32(ctx, "tokenizer.ggml.padding_token_id", ...);
gguf_set_val_u32(ctx, "llama.context_length", model->hparams.n_ctx);
gguf_set_val_u32(ctx, "llama.embedding_length", model->hparams.n_embd);
gguf_set_val_u32(ctx, "llama.feed_forward_length", model->hparams.n_ff);
gguf_set_val_u32(ctx, "llama.attention.head_count", model->hparams.n_head);
gguf_set_val_u32(ctx, "llama.block_count", model->hparams.n_layer);
gguf_set_val_u32(ctx, "llama.rope.dimension_count", model->hparams.n_rot);
gguf_set_val_f32(ctx, "llama.attention.layer_norm_rms_epsilon", 1e-5f);
// // n_head_kv is optional, default to n_head
// gguf_set_val_u32(ctx, "llama.attention.head_count_kv", hparams.n_head_kv);
// write tensors // write tensors
write_tensor(&file, model->tok_embeddings); ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD);
write_tensor(&file, model->norm); gguf_add_tensor(ctx, model->tok_embeddings);
write_tensor(&file, model->output); // ?
ggml_set_name(model->norm, TN_OUTPUT_NORM);
gguf_add_tensor(ctx, model->norm);
ggml_set_name(model->output, TN_OUTPUT);
gguf_add_tensor(ctx, model->output); // ?
for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
auto & layer = model->layers[i]; auto & layer = model->layers[i];
write_tensor(&file, layer.attention_norm); ggml_format_name(layer.attention_norm, TN_ATTN_NORM, i);
write_tensor(&file, layer.wq); gguf_add_tensor(ctx, layer.attention_norm);
write_tensor(&file, layer.wk);
write_tensor(&file, layer.wv); ggml_format_name(layer.wq, TN_ATTN_Q, i);
write_tensor(&file, layer.wo); gguf_add_tensor(ctx, layer.wq);
write_tensor(&file, layer.ffn_norm);
write_tensor(&file, layer.w1); ggml_format_name(layer.wk, TN_ATTN_K, i);
write_tensor(&file, layer.w2); gguf_add_tensor(ctx, layer.wk);
write_tensor(&file, layer.w3);
ggml_format_name(layer.wv, TN_ATTN_V, i);
gguf_add_tensor(ctx, layer.wv);
ggml_format_name(layer.wo, TN_ATTN_OUTPUT, i);
gguf_add_tensor(ctx, layer.wo);
ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i);
gguf_add_tensor(ctx, layer.ffn_norm);
ggml_format_name(layer.w1, TN_FFN_GATE, i);
gguf_add_tensor(ctx, layer.w1);
ggml_format_name(layer.w2, TN_FFN_DOWN, i);
gguf_add_tensor(ctx, layer.w2);
ggml_format_name(layer.w3, TN_FFN_UP, i);
gguf_add_tensor(ctx, layer.w3);
} }
gguf_write_to_file(ctx, filename, false);
gguf_free(ctx);
} }
struct train_params get_default_train_params() { struct train_params get_default_train_params() {
@ -840,6 +850,7 @@ int main(int argc, char ** argv) {
model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx); model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
model.hparams.n_ctx = params.n_ctx; model.hparams.n_ctx = params.n_ctx;
model.hparams.n_embd = config.dim; //params.n_embd; model.hparams.n_embd = config.dim; //params.n_embd;
model.hparams.n_ff = config.hidden_dim;
model.hparams.n_mult = 32;//params.n_mult; model.hparams.n_mult = 32;//params.n_mult;
model.hparams.n_head = config.n_heads; //params.n_head; model.hparams.n_head = config.n_heads; //params.n_head;
model.hparams.n_layer = config.n_layers; //params.n_layer; model.hparams.n_layer = config.n_layers; //params.n_layer;