Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 17:25:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into concedo_experimental

Browse Source 
# Conflicts:
#	.dockerignore
#	.github/workflows/build.yml
#	CMakeLists.txt
#	Makefile
#	README.md
#	flake.lock
#	flake.nix
#	tests/CMakeLists.txt
This commit is contained in:
Concedo
2023-08-28 14:19:05 +08:00
parent 9d5b4238e8 103cfafc77
commit 4b00916ac7
53 changed files with 4980 additions and 1910 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/CMakeLists.txt
+1
View File
@@ -25,6 +25,7 @@ else()
add_subdirectory(simple)
add_subdirectory(embd-input)
add_subdirectory(llama-bench)
add_subdirectory(beam_search)
if (LLAMA_METAL)
add_subdirectory(metal)
endif()
examples/beam_search/CMakeLists.txt
+8
View File
@@ -0,0 +1,8 @@
set(TARGET beam_search)
add_executable(${TARGET} beam_search.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
if(TARGET BUILD_INFO)
add_dependencies(${TARGET} BUILD_INFO)
endif()
examples/beam_search/beam_search.cpp
+188
View File
@@ -0,0 +1,188 @@
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include "common.h"
#include "llama.h"
#include "build-info.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
#include <signal.h>
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#define NOMINMAX
#include <windows.h>
#include <signal.h>
#endif
// Used for debugging to print out beam tokens.
struct ostream_beam_view {
llama_context * ctx;
llama_beam_view beam_view;
};
std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) {
os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
}
return os << ')';
}
// Put here anything you want back in beam_search_callback().
struct beam_search_callback_data {
llama_context * ctx;
std::vector<llama_token> response;
};
// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
// For example, eob can be flagged due to maximum token length, stop words, etc.
bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) {
return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
}
// Function matching type llama_beam_search_callback_fn_t.
// Custom callback example is called each time the beams lengths increase:
// * Show progress by printing ',' following by number of convergent beam tokens if any.
// * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
// This is also called when the stop condition is met.
// Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
// Mark beams as EOS as needed.
for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
llama_beam_view& beam_view = beams_state.beam_views[i];
if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) {
beam_view.eob = true;
}
}
printf(","); // Show progress
if (const size_t n = beams_state.common_prefix_length) {
callback_data.response.resize(callback_data.response.size() + n);
assert(0u < beams_state.n_beams);
const llama_token * tokens = beams_state.beam_views[0].tokens;
std::copy(tokens, tokens + n, callback_data.response.end() - n);
printf("%lu", n);
}
fflush(stdout);
#if 1 // DEBUG: print current beams for this iteration
std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n";
for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
std::cout << "beams["<<i<<"]: " << ostream_beam_view{callback_data.ctx,beams_state.beam_views[i]} << std::endl;
}
#endif
}
int main(int argc, char ** argv)
{
gpt_params params;
//params.n_gpu_layers = 200;
//---------------------------------
// Print help :
//---------------------------------
if ( argc < 2 || argv[1][0] == '-' )
{
printf( "Usage: %s MODEL_PATH [BEAM_WIDTH=2] [PROMPT]\n" , argv[0] );
return 1 ;
}
//---------------------------------
// Load parameters :
//---------------------------------
params.model = argv[1];
params.n_beams = 2 < argc ? std::stoi(argv[2]) : 2;
if ( argc > 3 )
{
params.prompt = argv[3];
}
if ( params.prompt.empty() )
{
params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n";
}
//---------------------------------
// Init LLM :
//---------------------------------
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params( params );
if ( model == NULL )
{
fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
return 1;
}
//---------------------------------
// Tokenize the prompt :
//---------------------------------
std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
const size_t max_context_size = llama_n_ctx( ctx );
const size_t max_tokens_list_size = max_context_size - 4 ;
if (tokens_list.size() > max_tokens_list_size)
{
fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" ,
__func__ , tokens_list.size() , max_tokens_list_size );
return 1;
}
fprintf( stderr, "\n\n" );
// Print the tokens from the prompt :
for( auto id : tokens_list )
{
std::cout << llama_token_to_piece(ctx, id);
}
std::cout << std::flush;
int n_past = llama_get_kv_cache_token_count(ctx);
if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
{
fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
return 1;
}
n_past += tokens_list.size();
beam_search_callback_data callback_data{ctx, {}};
size_t const beam_width = static_cast<size_t>(params.n_beams);
int const n_predict = 256;
llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
std::cout << "\n\n";
for (llama_token const token_id : callback_data.response) {
std::cout << llama_token_to_piece(ctx,token_id);
}
std::cout << std::endl;
llama_free( ctx );
llama_free_model( model );
llama_backend_free();
return 0;
}
examples/convert-llama2c-to-ggml/README.md
+2 -6
View File
@@ -12,18 +12,14 @@ usage: ./convert-llama2c-to-ggml [options]
options:
-h, --help show this help message and exit
--copy-vocab-from-model FNAME model path from which to copy vocab (default 'tokenizer.bin')
--copy-vocab-from-model FNAME path of gguf llama model or llama2.c vocabulary from which to copy vocab (default 'models/7B/ggml-model-f16.gguf')
--llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model
--llama2c-output-model FNAME model path to save the converted llama2.c model (default ak_llama_model.bin')
```
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`
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`
`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
Now you can use the model with a command like:
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+218 -122
View File
@@ -10,9 +10,48 @@
#include <ctime>
#include <random>
#include <stdexcept>
#include <sstream>
#include <algorithm>
#include <string>
// GGUF keys & tensor names.
#define KV_GENERAL_ARCHITECTURE "general.architecture"
#define KV_GENERAL_NAME "general.name"
#define KV_TOKENIZER_MODEL "tokenizer.ggml.model"
#define KV_TOKENIZER_LIST "tokenizer.ggml.tokens"
#define KV_TOKENIZER_TOKEN_TYPE "tokenizer.ggml.token_type"
#define KV_TOKENIZER_SCORES "tokenizer.ggml.scores"
#define KV_TOKENIZER_BOS_ID "tokenizer.ggml.bos_token_id"
#define KV_TOKENIZER_EOS_ID "tokenizer.ggml.eos_token_id"
#define KV_TOKENIZER_UNK_ID "tokenizer.ggml.unknown_token_id"
#define KV_TOKENIZER_SEP_ID "tokenizer.ggml.seperator_token_id"
#define KV_TOKENIZER_PAD_ID "tokenizer.ggml.padding_token_id"
#define KV_TOKENIZER_HF_JSON "tokenizer.huggingface.json"
#define KV_CONTEXT_LENGTH "llama.context_length"
#define KV_EMBEDDING_LENGTH "llama.embedding_length"
#define KV_BLOCK_COUNT "llama.block_count"
#define KV_FEED_FORWARD_LENGTH "llama.feed_forward_length"
#define KV_ATTENTION_HEAD_COUNT "llama.attention.head_count"
#define KV_ATTENTION_HEAD_COUNT_KV "llama.attention.head_count_kv"
#define KV_ATTENTION_LAYERNORM_RMS_EPS "llama.attention.layer_norm_rms_epsilon"
#define KV_ROPE_DIMENSION_COUNT "llama.rope.dimension_count"
#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)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
@@ -20,6 +59,11 @@
#define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt'
#define LLAMA_FILE_VERSION_GGJT_V3 3
#define TOKENIZER_NAME "llama"
#define UNKNOWN_TOKEN_ID 0
#define BOS_TOKEN_ID 1
#define EOS_TOKEN_ID 2
//////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
typedef struct {
int dim; // transformer dimension
@@ -183,6 +227,7 @@ struct my_llama_hparams {
uint32_t n_vocab = 32000;
uint32_t n_ctx = 512; // this is provided as user input?
uint32_t n_embd = 4096;
uint32_t n_ff = 11008;
uint32_t n_mult = 4;
uint32_t n_head = 32;
uint32_t n_layer = 32;
@@ -214,6 +259,8 @@ struct my_llama_layer {
struct my_llama_model {
struct ggml_context * ctx = NULL;
std::string name;
my_llama_hparams hparams;
struct ggml_tensor * tok_embeddings;
@@ -276,18 +323,13 @@ struct train_params {
int mem_compute1_gb;
};
uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
return n_ff;
}
void print_params(struct my_llama_hparams * params) {
printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
printf("%s: n_ctx: %d\n", __func__, params->n_ctx);
printf("%s: n_embd: %d\n", __func__, params->n_embd);
printf("%s: n_mult: %d\n", __func__, params->n_mult);
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_rot: %d\n", __func__, params->n_rot);
}
@@ -299,7 +341,7 @@ void init_model(struct my_llama_model * model) {
const uint32_t n_layer = hparams.n_layer;
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;
model->train_its = 0;
@@ -481,21 +523,6 @@ struct llama_file {
return std::string(chars.data(), len);
}
void write_raw(const void * ptr, size_t size) {
if (size == 0) {
return;
}
errno = 0;
size_t ret = std::fwrite(ptr, size, 1, fp);
if (ret != 1) {
throw std::runtime_error(format("write error: %s", strerror(errno)));
}
}
void write_u32(std::uint32_t val) {
write_raw(&val, sizeof(val));
}
~llama_file() {
if (fp) {
std::fclose(fp);
@@ -503,30 +530,6 @@ 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) {
llama_file file(filename, "rb");
if (file.size < 4) {
@@ -536,48 +539,96 @@ bool is_ggml_file(const char *filename) {
return magic == GGUF_MAGIC;
}
static std::string llama_escape_whitespaces(const std::string& text) {
std::ostringstream out;
for (char c : text) {
if (c == ' ') out << "\xe2\x96\x81";
else out << c;
}
return out.str();
}
void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
#pragma message("TODO: implement reading vocabulary using gguf")
// // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
// if (is_ggml_file(filename)) {
//
// struct llama_context_params llama_params = llama_context_default_params();
// llama_params.vocab_only = true;
//
// struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
// struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
//
// const int n_vocab = llama_n_vocab(lctx);
// vocab->id_to_token.resize(n_vocab);
// for (int i=0; i<n_vocab; ++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);
if (is_ggml_file(filename)) {
struct ggml_context * ctx_data = NULL;
struct gguf_init_params params = {
/*.no_alloc = */ false,
/*.ctx = */ &ctx_data,
};
struct gguf_context * ctx = gguf_init_from_file(filename, params);
GGML_ASSERT(ctx != NULL);
const int model_idx = gguf_find_key(ctx, KV_TOKENIZER_MODEL);
GGML_ASSERT(model_idx >= 0);
std::string tokenizer_name = gguf_get_val_str(ctx, model_idx);
GGML_ASSERT(tokenizer_name == TOKENIZER_NAME);
const int token_idx = gguf_find_key(ctx, KV_TOKENIZER_LIST);
GGML_ASSERT(token_idx >= 0);
const int score_idx = gguf_find_key(ctx, KV_TOKENIZER_SCORES);
GGML_ASSERT(score_idx >= 0);
const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
const int toktype_idx = gguf_find_key(ctx, KV_TOKENIZER_TOKEN_TYPE);
GGML_ASSERT(toktype_idx >= 0);
const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
vocab->id_to_token.resize(n_vocab);
for (uint32_t i = 0; i < n_vocab; i++) {
std::string word = gguf_get_arr_str(ctx, token_idx, i);
vocab->token_to_id[word] = i;
auto & token_data = vocab->id_to_token[i];
token_data.text = std::move(word);
token_data.score = scores[i];
token_data.type = (llama_token_type) toktypes[i];
}
ggml_free(ctx_data);
gguf_free(ctx);
} else {
// assume llama2.c vocabulary
printf("Assuming llama2.c vocabulary since %s is not a gguf 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) {
for (llama_vocab::id id=0; id<n_vocab; ++id) {
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;
unsigned char byte_val;
llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
if (id == UNKNOWN_TOKEN_ID) {
text = "<unk>";
type = LLAMA_TOKEN_TYPE_UNKNOWN;
} else if (id == BOS_TOKEN_ID) {
text = "<s>";
type = LLAMA_TOKEN_TYPE_CONTROL;
} else if (id == EOS_TOKEN_ID) {
text = "</s>";
type = LLAMA_TOKEN_TYPE_CONTROL;
} else if (text.empty()) {
type = LLAMA_TOKEN_TYPE_CONTROL;
} else if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
// Text of byte tokens is already in the expected format.
type = LLAMA_TOKEN_TYPE_BYTE;
} else {
type = LLAMA_TOKEN_TYPE_NORMAL;
}
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);
text = llama_escape_whitespaces(text);
vocab->id_to_token[id].text = text;
vocab->id_to_token[id].score = score;
vocab->id_to_token[id].type = type;
vocab->token_to_id.emplace(text, id);
}
}
}
@@ -619,33 +670,6 @@ 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) {
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.
// w->token_embedding_table -> model->tok_embeddings
// float* -> struct ggml_tensor
@@ -658,8 +682,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
// 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);
int n_ff = model->hparams.n_ff;
for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
auto & layer = model->layers[i];
@@ -677,28 +700,91 @@ 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.w3 , &w->w3[i*row_length*n_ff]);
}
struct gguf_context * ctx = gguf_init_empty();
std::vector<const char*> tokens;
std::vector<float> scores;
std::vector<llama_token_type> token_types;
for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
tokens.push_back(token_data.text.c_str());
scores.push_back(token_data.score);
token_types.push_back(token_data.type);
}
gguf_set_arr_str(ctx, KV_TOKENIZER_LIST, tokens.data(), tokens.size());
gguf_set_arr_data(ctx, KV_TOKENIZER_SCORES, GGUF_TYPE_FLOAT32, scores.data(), scores.size());
gguf_set_arr_data(ctx, KV_TOKENIZER_TOKEN_TYPE, GGUF_TYPE_INT32, token_types.data(), token_types.size());
gguf_set_val_str(ctx, KV_TOKENIZER_MODEL, TOKENIZER_NAME);
gguf_set_val_str(ctx, KV_GENERAL_ARCHITECTURE, "llama");
gguf_set_val_str(ctx, KV_GENERAL_NAME, "llama");
// special tokens
gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
// n_head_kv is optional, default to n_head
// gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
// write tensors
write_tensor(&file, model->tok_embeddings);
write_tensor(&file, model->norm);
write_tensor(&file, model->output); // ?
ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD);
gguf_add_tensor(ctx, model->tok_embeddings);
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) {
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);
ggml_format_name(layer.wq, TN_ATTN_Q, i);
gguf_add_tensor(ctx, layer.wq);
ggml_format_name(layer.wk, TN_ATTN_K, i);
gguf_add_tensor(ctx, layer.wk);
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.attention_norm, TN_ATTN_NORM, i);
gguf_add_tensor(ctx, layer.attention_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);
ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i);
gguf_add_tensor(ctx, layer.ffn_norm);
}
gguf_write_to_file(ctx, filename, false);
gguf_free(ctx);
}
struct train_params get_default_train_params() {
struct train_params params;
params.fn_vocab_model = "tokenizer.bin";
params.fn_vocab_model = "models/7B/ggml-model-f16.gguf";
params.fn_llama2c_output_model = "ak_llama_model.bin";
params.fn_train_data = "shakespeare.txt";
params.fn_checkpoint_in = "checkpoint.bin";
@@ -751,7 +837,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " --copy-vocab-from-model FNAME llama2.c vocabulary or ggmlv3 model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
fprintf(stderr, " --copy-vocab-from-model FNAME path of gguf llama model or llama2.c vocabulary from which to copy vocab (default '%s')\n", params->fn_vocab_model);
fprintf(stderr, " --llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
fprintf(stderr, " --llama2c-output-model FNAME model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
fprintf(stderr, "\n");
@@ -812,6 +898,14 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
return true;
}
std::string basename(const std::string &path) {
size_t pos = path.find_last_of("/");
if (pos == std::string::npos) {
return path;
}
return path.substr(pos + 1);
}
int main(int argc, char ** argv) {
struct train_params params = get_default_train_params();
if (!params_parse(argc, argv, &params)) {
@@ -840,6 +934,7 @@ int main(int argc, char ** argv) {
model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
model.hparams.n_ctx = params.n_ctx;
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_head = config.n_heads; //params.n_head;
model.hparams.n_layer = config.n_layers; //params.n_layer;
@@ -853,6 +948,7 @@ int main(int argc, char ** argv) {
model.ctx = ggml_init(lcparams);
init_model(&model);
model.name = basename(params.fn_llama2c_model);
save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
examples/embd-input/embd-input-lib.cpp
+1 -1
View File
@@ -214,7 +214,7 @@ const char * sampling(struct MyModel * mymodel) {
if (id == llama_token_eos(ctx)) {
ret = "</s>";
} else {
ret = llama_token_to_str(ctx, id);
ret = llama_token_to_piece(ctx, id);
}
eval_id(mymodel, id);
return ret.c_str();
examples/embedding/embedding.cpp
+1 -4
View File
@@ -56,9 +56,6 @@ int main(int argc, char ** argv) {
int n_past = 0;
// Add a space in front of the first character to match OG llama tokenizer behavior
params.prompt.insert(0, 1, ' ');
// tokenize the prompt
auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
@@ -67,7 +64,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
fprintf(stderr, "\n");
}
examples/gguf/gguf.cpp
+3
View File
@@ -30,6 +30,9 @@ bool gguf_ex_write(const std::string & fname) {
gguf_set_val_u32 (ctx, "some.parameter.uint32", 0x12345678);
gguf_set_val_i32 (ctx, "some.parameter.int32", -0x12345679);
gguf_set_val_f32 (ctx, "some.parameter.float32", 0.123456789f);
gguf_set_val_u64 (ctx, "some.parameter.uint64", 0x123456789abcdef0ull);
gguf_set_val_i64 (ctx, "some.parameter.int64", -0x123456789abcdef1ll);
gguf_set_val_f64 (ctx, "some.parameter.float64", 0.1234567890123456789);
gguf_set_val_bool(ctx, "some.parameter.bool", true);
gguf_set_val_str (ctx, "some.parameter.string", "hello world");
examples/llama-bench/llama-bench.cpp
+53 -11
View File
@@ -18,9 +18,7 @@
#include "llama.h"
#include "common.h"
#include "build-info.h"
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
#endif
// utils
static uint64_t get_time_ns() {
@@ -443,6 +441,8 @@ struct test {
static const std::string gpu_info;
std::string model_filename;
std::string model_type;
uint64_t model_size;
uint64_t model_n_params;
int n_batch;
int n_threads;
bool f32_kv;
@@ -459,8 +459,10 @@ struct test {
test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
model_filename = inst.model;
char buf[128];
llama_model_type(lmodel, buf, sizeof(buf));
llama_model_desc(lmodel, buf, sizeof(buf));
model_type = buf;
model_size = llama_model_size(lmodel);
model_n_params = llama_model_n_params(lmodel);
n_batch = inst.n_batch;
n_threads = inst.n_threads;
f32_kv = inst.f32_kv;
@@ -504,7 +506,7 @@ struct test {
static std::string get_backend() {
if (cuda) {
return "CUDA";
return GGML_CUDA_NAME;
}
if (opencl) {
return "OpenCL";
@@ -526,7 +528,7 @@ struct test {
"build_commit", "build_number",
"cuda", "opencl", "metal", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"model_filename", "model_type",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_threads", "f16_kv",
"n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
"n_prompt", "n_gen", "test_time",
@@ -540,6 +542,7 @@ struct test {
static field_type get_field_type(const std::string & field) {
if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
field == "model_size" || field == "model_n_params" ||
field == "n_gpu_layers" || field == "main_gpu" ||
field == "n_prompt" || field == "n_gen" ||
field == "avg_ns" || field == "stddev_ns") {
@@ -575,7 +578,7 @@ struct test {
build_commit, std::to_string(build_number),
std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
model_filename, model_type,
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
std::to_string(n_prompt), std::to_string(n_gen), test_time,
@@ -711,8 +714,15 @@ struct markdown_printer : public printer {
return -30;
}
if (field == "t/s") {
return 15;
return 16;
}
if (field == "size" || field == "params") {
return 10;
}
if (field == "n_gpu_layers") {
return 3;
}
int width = std::max((int)field.length(), 10);
if (test::get_field_type(field) == test::STRING) {
@@ -721,9 +731,28 @@ struct markdown_printer : public printer {
return width;
}
static std::string get_field_display_name(const std::string & field) {
if (field == "n_gpu_layers") {
return "ngl";
}
if (field == "n_threads") {
return "threads";
}
if (field == "mul_mat_q") {
return "mmq";
}
if (field == "tensor_split") {
return "ts";
}
return field;
}
void print_header(const cmd_params & params) override {
// select fields to print
fields = { "model", "backend" };
fields.push_back("model");
fields.push_back("size");
fields.push_back("params");
fields.push_back("backend");
bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
if (!is_cpu_backend) {
fields.push_back("n_gpu_layers");
@@ -754,7 +783,7 @@ struct markdown_printer : public printer {
fprintf(fout, "|");
for (const auto & field : fields) {
fprintf(fout, " %*s |", get_field_width(field), field.c_str());
fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str());
}
fprintf(fout, "\n");
fprintf(fout, "|");
@@ -771,12 +800,26 @@ struct markdown_printer : public printer {
fprintf(fout, "|");
for (const auto & field : fields) {
std::string value;
char buf[128];
if (field == "model") {
value = t.model_type;
} else if (field == "size") {
if (t.model_size < 1024*1024*1024) {
snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
} else {
snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
}
value = buf;
} else if (field == "params") {
if (t.model_n_params < 1000*1000*1000) {
snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
} else {
snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
}
value = buf;
} else if (field == "backend") {
value = test::get_backend();
} else if (field == "test") {
char buf[128];
if (t.n_prompt > 0 && t.n_gen == 0) {
snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
} else if (t.n_gen > 0 && t.n_prompt == 0) {
@@ -787,7 +830,6 @@ struct markdown_printer : public printer {
}
value = buf;
} else if (field == "t/s") {
char buf[128];
snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
value = buf;
} else if (vmap.find(field) != vmap.end()) {
examples/main/main.cpp
+22 -15
View File
@@ -189,12 +189,14 @@ int main(int argc, char ** argv) {
}
}
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
// Add BOS if SPM tokenizer
const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
// tokenize the prompt
std::vector<llama_token> embd_inp;
if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
} else {
embd_inp = session_tokens;
}
@@ -209,10 +211,9 @@ int main(int argc, char ** argv) {
int guidance_offset = 0;
int original_prompt_len = 0;
if (ctx_guidance) {
params.cfg_negative_prompt.insert(0, 1, ' ');
guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm);
guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
}
@@ -259,7 +260,7 @@ int main(int argc, char ** argv) {
}
// prefix & suffix for instruct mode
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm);
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
// in instruct mode, we inject a prefix and a suffix to each input by the user
@@ -278,7 +279,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (ctx_guidance) {
@@ -286,14 +287,14 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
for (int i = 0; i < (int) guidance_inp.size(); i++) {
fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_str(ctx, guidance_inp[i]).c_str());
fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
}
}
if (params.n_keep > 0) {
fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
fprintf(stderr, "%s", llama_token_to_str(ctx, embd_inp[i]).c_str());
fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
fprintf(stderr, "'\n");
}
@@ -449,7 +450,7 @@ int main(int argc, char ** argv) {
//printf("\n---\n");
//printf("resetting: '");
//for (int i = 0; i < (int) embd.size(); i++) {
// printf("%s", llama_token_to_str(ctx, embd[i]));
// printf("%s", llama_token_to_piece(ctx, embd[i]));
//}
//printf("'\n");
//printf("\n---\n");
@@ -502,7 +503,7 @@ int main(int argc, char ** argv) {
input_size = embd_guidance.size();
//fprintf(stderr, "\n---------------------\n");
//for (int i = 0; i < (int) embd_guidance.size(); i++) {
//fprintf(stderr, "%s", llama_token_to_str(ctx, embd_guidance[i]));
//fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i]));
//}
//fprintf(stderr, "\n---------------------\n");
} else {
@@ -597,7 +598,12 @@ int main(int argc, char ** argv) {
last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
last_n_repeat, alpha_frequency, alpha_presence);
if (!penalize_nl) {
logits[llama_token_nl(ctx)] = nl_logit;
for (size_t idx = 0; idx < candidates_p.size; idx++) {
if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
candidates_p.data[idx].logit = nl_logit;
break;
}
}
}
if (grammar != NULL) {
@@ -661,7 +667,7 @@ int main(int argc, char ** argv) {
// display text
if (input_echo) {
for (auto id : embd) {
printf("%s", llama_token_to_str(ctx, id).c_str());
printf("%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stdout);
}
@@ -677,7 +683,7 @@ int main(int argc, char ** argv) {
if (params.antiprompt.size()) {
std::string last_output;
for (auto id : last_n_tokens) {
last_output += llama_token_to_str(ctx, id);
last_output += llama_token_to_piece(ctx, id);
}
is_antiprompt = false;
@@ -798,7 +804,8 @@ int main(int argc, char ** argv) {
}
// In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
if (params.interactive && n_remain <= 0 && params.n_predict != -1) {
// We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
n_remain = params.n_predict;
is_interacting = true;
}
examples/perplexity/perplexity.cpp
+78 -15
View File
@@ -6,6 +6,8 @@
#include <ctime>
#include <sstream>
#include <cstring>
#include <thread>
#include <mutex>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
@@ -27,6 +29,40 @@ std::vector<float> softmax(const std::vector<float>& logits) {
return probs;
}
float log_softmax(int n_vocab, const float * logits, int tok) {
float max_logit = logits[0];
for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
double sum_exp = 0.0;
for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
return logits[tok] - max_logit - log(sum_exp);
}
void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread>& workers,
double& nll, double& nll2) {
std::mutex mutex;
int counter = 0;
auto compute = [&mutex, &counter, &nll, &nll2, n_vocab, logits, tokens, n_token] () {
double local_nll = 0, local_nll2 = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
int i = counter++;
if (i >= n_token) {
nll += local_nll; nll2 += local_nll2;
break;
}
lock.unlock();
double v = -log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
local_nll += v;
local_nll2 += v*v;
}
};
for (auto& w : workers) w = std::thread(compute);
compute();
for (auto& w : workers) w.join();
}
void perplexity_v2(llama_context * ctx, const gpt_params & params) {
// Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
@@ -154,10 +190,14 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = is_spm;
auto tim1 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
auto tim2 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
const int n_chunk_max = tokens.size() / params.n_ctx;
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
@@ -166,9 +206,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
int count = 0;
double nll = 0.0;
double nll2 = 0.0;
fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
for (int i = 0; i < n_chunk; ++i) {
const int start = i * params.n_ctx;
const int end = start + params.n_ctx;
@@ -228,26 +271,32 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
// Example, we have a context window of 512, we will compute perplexity for each of the
// last 256 tokens. Then, we split the input up into context window size chunks to
// process the entire prompt.
for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
// Calculate probability of next token, given the previous ones.
const std::vector<float> tok_logits(
logits.begin() + (j + 0) * n_vocab,
logits.begin() + (j + 1) * n_vocab);
const int first = std::min(512, params.n_ctx/2);
process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, workers, nll, nll2);
count += params.n_ctx - first - 1;
const float prob = softmax(tok_logits)[tokens[start + j + 1]];
nll += -std::log(prob);
++count;
}
// perplexity is e^(average negative log-likelihood)
if (params.ppl_output_type == 0) {
printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
} else {
printf("%8d %.4lf\n", i*params.n_ctx, std::exp(nll / count));
double av = nll/count;
double av2 = nll2/count - av*av;
if (av2 > 0) av2 = sqrt(av2/(count-1));
printf("%8d %.4lf %4lf %4lf\n", i*params.n_ctx, std::exp(nll / count), av, av2);
}
fflush(stdout);
}
printf("\n");
nll2 /= count;
nll /= count;
nll2 -= nll * nll;
if (nll2 > 0) {
nll2 = sqrt(nll2/(count-1));
double ppl = exp(nll);
printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
} else {
printf("Unexpected negative standard deviation of log(prob)\n");
}
}
std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
@@ -306,6 +355,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
fprintf(stderr, "================================= is_spm = %d\n", is_spm);
// This is needed as usual for LLaMA models
const bool add_bos = is_spm;
@@ -346,7 +396,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
hs_data[i].context = prompt_lines[idx*6];
hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
for (size_t j=0; j < 4; j++) {
hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
hs_data[i].ending[j] = prompt_lines[idx*6+2+j];
}
// Delete the selected random example from the prompt
@@ -361,6 +411,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
double acc = 0.0f;
const int n_vocab = llama_n_vocab(ctx);
std::vector<std::vector<int>> ending_tokens(4);
std::vector<float> tok_logits(n_vocab);
for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
@@ -368,11 +420,21 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
size_t context_size = context_embd.size();
for (int i = 0; i < 4; ++i) {
ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + " " + hs_data[task_idx].ending[i], add_bos);
for (int k = 0; k < int(context_size); ++k) {
if (ending_tokens[i][k] != context_embd[k]) {
fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k);
break;
}
}
}
// Do the 1st ending
// In this case we include the context when evaluating
auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
//auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
auto query_embd = ending_tokens[0];
auto query_size = query_embd.size();
//printf("First query: %d\n",(int)query_size);
// Stop if query wont fit the ctx window
if (query_size > (size_t)params.n_ctx) {
@@ -417,7 +479,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) {
// Tokenize the query
query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false);
query_embd.resize(ending_tokens[ending_idx].size() - context_size);
std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int));
query_size = query_embd.size();
// Stop if query wont fit the ctx window
examples/save-load-state/save-load-state.cpp
+2 -2
View File
@@ -87,7 +87,7 @@ int main(int argc, char ** argv) {
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
auto next_token = llama_sample_token(ctx, &candidates_p);
auto next_token_str = llama_token_to_str(ctx, next_token);
auto next_token_str = llama_token_to_piece(ctx, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str.c_str());
@@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
auto next_token = llama_sample_token(ctx2, &candidates_p);
auto next_token_str = llama_token_to_str(ctx2, next_token);
auto next_token_str = llama_token_to_piece(ctx2, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str.c_str());
examples/server/README.md
+16 -13
View File
@@ -77,34 +77,31 @@ You need to have [Node.js](https://nodejs.org/en) installed.
```bash
mkdir llama-client
cd llama-client
npm init
npm install axios
```
Create a index.js file and put inside this:
```javascript
const axios = require("axios");
const prompt = `Building a website can be done in 10 simple steps:`;
async function Test() {
let result = await axios.post("http://127.0.0.1:8080/completion", {
prompt,
n_predict: 512,
});
// the response is received until completion finish
console.log(result.data.content);
let response = await fetch("http://127.0.0.1:8080/completion", {
method: 'POST',
body: JSON.stringify({
prompt,
n_predict: 512,
})
})
console.log((await response.json()).content)
}
Test();
Test()
```
And run it:
```bash
node .
node index.js
```
## API Endpoints
@@ -167,6 +164,12 @@ node .
Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
- **POST** `/detokenize`: Convert tokens to text.
*Options:*
`tokens`: Set the tokens to detokenize.
- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
*Options:*
examples/server/index.html.hpp
+1691 -1117
View File
File diff suppressed because it is too large Load Diff
examples/server/public/index.html
+231 -14
View File
@@ -102,6 +102,17 @@
padding: 0.5em;
}
.prob-set {
padding: 0.3em;
border-bottom: 1px solid #ccc;
}
.popover-content {
position: absolute;
background-color: white;
padding: 0.2em;
box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);
}
textarea {
padding: 5px;
@@ -133,11 +144,17 @@
font-size: 80%;
color: #888;
}
@media (prefers-color-scheme: dark) {
.popover-content {
background-color: black;
}
}
</style>
<script type="module">
import {
html, h, signal, effect, computed, render, useSignal, useEffect, useRef
html, h, signal, effect, computed, render, useSignal, useEffect, useRef, Component
} from '/index.js';
import { llama } from '/completion.js';
@@ -168,6 +185,7 @@
mirostat_tau: 5, // target entropy
mirostat_eta: 0.1, // learning rate
grammar: '',
n_probs: 0, // no completion_probabilities
})
/* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -334,10 +352,21 @@
const prompt = template(session.value.template, {
message: msg,
history: session.value.transcript.flatMap(([name, message]) => template(session.value.historyTemplate, {name, message})).join("\n"),
history: session.value.transcript.flatMap(
([name, data]) =>
template(
session.value.historyTemplate,
{
name,
message: Array.isArray(data) ?
data.map(msg => msg.content).join('').replace(/^\s/, '') :
data,
}
)
).join("\n"),
});
let currentMessage = '';
const currentMessages = [];
const history = session.value.transcript
const llamaParams = {
@@ -347,15 +376,19 @@
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
const data = chunk.data;
currentMessage += data.content;
// remove leading whitespace
currentMessage = currentMessage.replace(/^\s+/, "")
transcriptUpdate([...history, ["{{char}}", currentMessage]])
if (data.stop) {
console.log("Completion finished: '", currentMessage, "', summary: ", data);
while (
currentMessages.length > 0 &&
currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
) {
currentMessages.pop();
}
transcriptUpdate([...history, ["{{char}}", currentMessages]])
console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
} else {
currentMessages.push(data);
transcriptUpdate([...history, ["{{char}}", currentMessages]])
}
if (data.timings) {
@@ -420,8 +453,18 @@
}
}, [messages])
const chatLine = ([user, msg]) => {
return html`<p key=${msg}><strong>${template(user)}:</strong> <${Markdownish} text=${template(msg)} /></p>`
const chatLine = ([user, data], index) => {
let message
const isArrayMessage = Array.isArray(data)
if (params.value.n_probs > 0 && isArrayMessage) {
message = html`<${Probabilities} data=${data} />`
} else {
const text = isArrayMessage ?
data.map(msg => msg.content).join('').replace(/^\s+/, '') :
data;
message = html`<${Markdownish} text=${template(text)} />`
}
return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
};
return html`
@@ -568,10 +611,71 @@
${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
</fieldset>
<fieldset>
${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
</fieldset>
</details>
</form>
`
}
const probColor = (p) => {
const r = Math.floor(192 * (1 - p));
const g = Math.floor(192 * p);
return `rgba(${r},${g},0,0.3)`;
}
const Probabilities = (params) => {
return params.data.map(msg => {
const { completion_probabilities } = msg;
if (
!completion_probabilities ||
completion_probabilities.length === 0
) return msg.content
if (completion_probabilities.length > 1) {
// Not for byte pair
if (completion_probabilities[0].content.startsWith('byte: \\')) return msg.content
const splitData = completion_probabilities.map(prob => ({
content: prob.content,
completion_probabilities: [prob]
}))
return html`<${Probabilities} data=${splitData} />`
}
const { probs, content } = completion_probabilities[0]
const found = probs.find(p => p.tok_str === msg.content)
const pColor = found ? probColor(found.prob) : 'transparent'
const popoverChildren = html`
<div class="prob-set">
${probs.map((p, index) => {
return html`
<div
key=${index}
title=${`prob: ${p.prob}`}
style=${{
padding: '0.3em',
backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
}}
>
<span>${p.tok_str}: </span>
<span>${Math.floor(p.prob * 100)}%</span>
</div>
`
})}
</div>
`
return html`
<${Popover} style=${{ backgroundColor: pColor }} popoverChildren=${popoverChildren}>
${msg.content.match(/\n/gim) ? html`<br />` : msg.content}
</>
`
});
}
// poor mans markdown replacement
const Markdownish = (params) => {
const md = params.text
@@ -600,10 +704,121 @@
`
}
// simple popover impl
const Popover = (props) => {
const isOpen = useSignal(false);
const position = useSignal({ top: '0px', left: '0px' });
const buttonRef = useRef(null);
const popoverRef = useRef(null);
const togglePopover = () => {
if (buttonRef.current) {
const rect = buttonRef.current.getBoundingClientRect();
position.value = {
top: `${rect.bottom + window.scrollY}px`,
left: `${rect.left + window.scrollX}px`,
};
}
isOpen.value = !isOpen.value;
};
const handleClickOutside = (event) => {
if (popoverRef.current && !popoverRef.current.contains(event.target) && !buttonRef.current.contains(event.target)) {
isOpen.value = false;
}
};
useEffect(() => {
document.addEventListener('mousedown', handleClickOutside);
return () => {
document.removeEventListener('mousedown', handleClickOutside);
};
}, []);
return html`
<span style=${props.style} ref=${buttonRef} onClick=${togglePopover}>${props.children}</span>
${isOpen.value && html`
<${Portal} into="#portal">
<div
ref=${popoverRef}
class="popover-content"
style=${{
top: position.value.top,
left: position.value.left,
}}
>
${props.popoverChildren}
</div>
</${Portal}>
`}
`;
};
// Source: preact-portal (https://github.com/developit/preact-portal/blob/master/src/preact-portal.js)
/** Redirect rendering of descendants into the given CSS selector */
class Portal extends Component {
componentDidUpdate(props) {
for (let i in props) {
if (props[i] !== this.props[i]) {
return setTimeout(this.renderLayer);
}
}
}
componentDidMount() {
this.isMounted = true;
this.renderLayer = this.renderLayer.bind(this);
this.renderLayer();
}
componentWillUnmount() {
this.renderLayer(false);
this.isMounted = false;
if (this.remote && this.remote.parentNode) this.remote.parentNode.removeChild(this.remote);
}
findNode(node) {
return typeof node === 'string' ? document.querySelector(node) : node;
}
renderLayer(show = true) {
if (!this.isMounted) return;
// clean up old node if moving bases:
if (this.props.into !== this.intoPointer) {
this.intoPointer = this.props.into;
if (this.into && this.remote) {
this.remote = render(html`<${PortalProxy} />`, this.into, this.remote);
}
this.into = this.findNode(this.props.into);
}
this.remote = render(html`
<${PortalProxy} context=${this.context}>
${show && this.props.children || null}
</${PortalProxy}>
`, this.into, this.remote);
}
render() {
return null;
}
}
// high-order component that renders its first child if it exists.
// used as a conditional rendering proxy.
class PortalProxy extends Component {
getChildContext() {
return this.props.context;
}
render({ children }) {
return children || null;
}
}
function App(props) {
return html`
<div id="container">
<div>
<header>
<h1>llama.cpp</h1>
</header>
@@ -624,11 +839,13 @@
`;
}
render(h(App), document.body);
render(h(App), document.querySelector('#container'));
</script>
</head>
<body>
<div id="container"></div>
<div id="portal"></div>
</body>
</html>
examples/server/server.cpp
+159 -48
View File
@@ -94,7 +94,7 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
std::string ret;
for (; begin != end; ++begin)
{
ret += llama_token_to_str(ctx, *begin);
ret += llama_token_to_piece(ctx, *begin);
}
return ret;
}
@@ -123,9 +123,10 @@ static void server_log(const char *level, const char *function, int line,
// format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
{
std::string out = token == -1 ? "" : llama_token_to_str(ctx, token);
// if first bit is 1, meaning it's a partial character
if (out.size() > 0 && (out[0] & 0x80) == 0x80)
std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
// if the size is 1 and first bit is 1, meaning it's a partial character
// (size > 1 meaning it's already a known token)
if (out.size() == 1 && (out[0] & 0x80) == 0x80)
{
std::stringstream ss;
ss << std::hex << (out[0] & 0xff);
@@ -285,7 +286,6 @@ struct llama_server_context
std::vector<llama_token> p;
if (first)
{
s.insert(0, 1, ' '); // add a space if it's the first
p = ::llama_tokenize(ctx, s, add_bos);
first = false;
}
@@ -308,7 +308,6 @@ struct llama_server_context
else
{
auto s = json_prompt.template get<std::string>();
s.insert(0, 1, ' '); // always add a first space
prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
}
@@ -565,7 +564,7 @@ struct llama_server_context
if (!embd.empty() && embd.back() == llama_token_eos(ctx))
{
// stopping_word = llama_token_to_str(ctx, embd.back());
// stopping_word = llama_token_to_piece(ctx, embd.back());
has_next_token = false;
stopped_eos = true;
LOG_VERBOSE("eos token found", {});
@@ -612,7 +611,7 @@ struct llama_server_context
{
const completion_token_output token_with_probs = nextToken();
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(ctx, token_with_probs.tok);
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
generated_text += token_text;
if (params.n_probs > 0)
@@ -1103,6 +1102,12 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
{"tokens", tokens}};
}
static json format_detokenized_response(std::string content)
{
return json{
{"content", content}};
}
template <typename T>
static T json_value(const json &body, const std::string &key, const T &default_value)
{
@@ -1208,6 +1213,62 @@ static void log_server_request(const Request &req, const Response &res)
});
}
bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
}
// Function matching type llama_beam_search_callback_fn_t.
// Custom callback example is called each time the beams lengths increase:
// * Show progress by printing ',' following by number of convergent beam tokens if any.
// * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
// This is also called when the stop condition is met.
// Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
auto & llama = *static_cast<llama_server_context*>(callback_data);
// Mark beams as EOS as needed.
for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
llama_beam_view& beam_view = beams_state.beam_views[i];
if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) {
beam_view.eob = true;
}
}
printf(","); // Show progress
if (const size_t n = beams_state.common_prefix_length) {
llama.generated_token_probs.resize(llama.generated_token_probs.size() + n);
assert(0u < beams_state.n_beams);
const llama_token * tokens = beams_state.beam_views[0].tokens;
const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
printf("%lu", n);
}
fflush(stdout);
#if 0 // DEBUG: print current beams for this iteration
std::cout << "\n\nCurrent beams:\n";
for (size_t i=0 ; i < beams_state.n_beams ; ++i) {
std::cout << "beams["<<i<<"]: " << ostream_beam_view{state.ctx,beams_state.beam_views[i]} << std::endl;
}
#endif
}
struct token_translator {
llama_context * ctx;
std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
std::string operator()(completion_token_output cto) const { return (*this)(cto.tok); }
};
void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
auto & gtps = llama.generated_token_probs;
auto translator = token_translator{llama.ctx};
auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
if (llama.generated_text.capacity() < llama.generated_text.size() + len) {
llama.generated_text.reserve(llama.generated_text.size() + len);
}
for (const completion_token_output & cto : gtps) {
llama.generated_text += translator(cto);
}
}
int main(int argc, char **argv)
{
// own arguments required by this example
@@ -1290,22 +1351,30 @@ int main(int argc, char **argv)
llama.beginCompletion();
if (!llama.stream) {
size_t stop_pos = std::string::npos;
if (llama.params.n_beams) {
// Fill llama.generated_token_probs vector with final beam.
llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
llama.n_past, llama.n_remain, llama.params.n_threads);
// Translate llama.generated_token_probs to llama.generated_text.
append_to_generated_text_from_generated_token_probs(llama);
} else {
size_t stop_pos = std::string::npos;
while (llama.has_next_token) {
const completion_token_output token_with_probs = llama.doCompletion();
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
while (llama.has_next_token) {
const completion_token_output token_with_probs = llama.doCompletion();
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok);
stop_pos = llama.findStoppingStrings(llama.generated_text,
token_text.size(), STOP_FULL);
}
stop_pos = llama.findStoppingStrings(llama.generated_text,
token_text.size(), STOP_FULL);
}
if (stop_pos == std::string::npos) {
stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
}
if (stop_pos != std::string::npos) {
llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
llama.generated_text.end());
if (stop_pos == std::string::npos) {
stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
}
if (stop_pos != std::string::npos) {
llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
llama.generated_text.end());
}
}
const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
@@ -1321,59 +1390,86 @@ int main(int argc, char **argv)
while (llama.has_next_token) {
const completion_token_output token_with_probs = llama.doCompletion();
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
if (llama.multibyte_pending > 0) {
if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
continue;
}
const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
size_t pos = std::min(sent_count, llama.generated_text.size());
const std::string str_test = llama.generated_text.substr(pos);
bool is_stop_full = false;
size_t stop_pos =
llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
if (stop_pos != std::string::npos) {
is_stop_full = true;
llama.generated_text.erase(
llama.generated_text.begin() + pos + stop_pos,
llama.generated_text.end());
pos = std::min(sent_count, llama.generated_text.size());
} else {
is_stop_full = false;
stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
STOP_PARTIAL);
}
const std::string to_send = llama.generated_text.substr(pos, stop_pos);
sent_count += to_send.size();
if (
stop_pos == std::string::npos ||
// Send rest of the text if we are at the end of the generation
(!llama.has_next_token && !is_stop_full && stop_pos > 0)
) {
const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
std::vector<completion_token_output> probs_output = {};
sent_count += to_send.size();
if (llama.params.n_probs > 0) {
const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
if (probs_pos < probs_stop_pos) {
probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
std::vector<completion_token_output> probs_output = {};
if (llama.params.n_probs > 0) {
const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
if (probs_pos < probs_stop_pos) {
probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
}
sent_token_probs_index = probs_stop_pos;
}
const json data = format_partial_response(llama, to_send, probs_output);
const std::string str =
"data: " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
if (!sink.write(str.data(), str.size())) {
LOG_VERBOSE("stream closed", {});
llama_print_timings(llama.ctx);
return false;
}
sent_token_probs_index = probs_stop_pos;
}
const json data = llama.has_next_token
? format_partial_response(llama, to_send, probs_output)
// Generation is done, send extra information.
: format_final_response(llama, to_send, llama.generated_token_probs);
if (!llama.has_next_token) {
// Generation is done, send extra information.
const json data = format_final_response(llama, "", llama.generated_token_probs);
const std::string str =
"data: " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
const std::string str =
"data: " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
if (!sink.write(str.data(), str.size())) {
LOG_VERBOSE("stream closed", {});
llama_print_timings(llama.ctx);
return false;
if (!sink.write(str.data(), str.size())) {
LOG_VERBOSE("stream closed", {});
llama_print_timings(llama.ctx);
return false;
}
}
}
@@ -1409,6 +1505,21 @@ int main(int argc, char **argv)
const json data = format_tokenizer_response(tokens);
return res.set_content(data.dump(), "application/json"); });
svr.Post("/detokenize", [&llama](const Request &req, Response &res)
{
auto lock = llama.lock();
const json body = json::parse(req.body);
std::string content;
if (body.count("tokens") != 0)
{
const std::vector<llama_token> tokens = body["tokens"];
content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
}
const json data = format_detokenized_response(content);
return res.set_content(data.dump(), "application/json"); });
svr.Post("/embedding", [&llama](const Request &req, Response &res)
{
auto lock = llama.lock();
examples/simple/simple.cpp
+2 -2
View File
@@ -63,7 +63,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "\n\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_str(ctx, id).c_str());
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
@@ -112,7 +112,7 @@ int main(int argc, char ** argv) {
}
// print the new token :
printf("%s", llama_token_to_str(ctx, new_token_id).c_str());
printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
// push this new token for next evaluation
examples/train-text-from-scratch/train-text-from-scratch.cpp
+2 -2
View File
@@ -1964,7 +1964,7 @@ void print_matrix(struct ggml_tensor * probs) {
void print_token(struct llama_context * ctx, llama_token token) {
printf("%s", llama_token_to_str(ctx, token).c_str());
printf("%s", llama_token_to_piece(ctx, token).c_str());
}
void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) {
@@ -2202,7 +2202,7 @@ int tokenize_file(struct llama_context * lctx, const char * filename, std::vecto
const char * in = buf.data();
const char * end = buf.data() + buf.size();
for (int i = 0; i < (int) out.size(); ++i) {
std::string s = llama_token_to_str(lctx, out[i]);
std::string s = llama_token_to_piece(lctx, out[i]);
int len = s.length();
if (in >= end) {
printf("%s: unexpected end of original text.\n", __func__);