Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into concedo_experimental

Browse Source 
# Conflicts:
#	.gitignore
#	README.md
#	tests/CMakeLists.txt
This commit is contained in:
Concedo
2023-08-24 15:21:24 +08:00
parent 8263fd7bdb 6e91a1b070
commit b8372d4466
51 changed files with 2120 additions and 825 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/convert-llama2c-to-ggml/README.md
+9 -5
View File
@@ -12,15 +12,19 @@ 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 'models/ggml-vocab.bin')
--copy-vocab-from-model FNAME model path from which to copy vocab (default 'tokenizer.bin')
--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 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 <ggml-vocab.bin> --llama2c-model <llama2.c model path> --llama2c-output-model <ggml output model path>`
`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin`
Now you can use the model with command like:
For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually:
`$ ./main -m <ggml output model path> -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5`
`$ 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:
`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+135 -99
View File
@@ -17,6 +17,9 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt'
#define LLAMA_FILE_VERSION_GGJT_V3 3
//////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
typedef struct {
int dim; // transformer dimension
@@ -49,10 +52,10 @@ typedef struct {
// float* freq_cis_real; // (seq_len, dim/2)
// float* freq_cis_imag; // (seq_len, dim/2)
// (optional) classifier weights for the logits, on the last layer
//float* wcls;
float* wcls;
} TransformerWeights;
void malloc_weights(TransformerWeights* w, Config* p) {
void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
// we calloc instead of malloc to keep valgrind happy
w->token_embedding_table = new float[p->vocab_size * p->dim]();
printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
@@ -86,9 +89,16 @@ void malloc_weights(TransformerWeights* w, Config* p) {
w->rms_final_weight = new float[p->dim]();
printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
if (shared_weights) {
w->wcls = NULL;
} else {
w->wcls = new float[p->vocab_size * p->dim]();
printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
}
}
int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
@@ -100,6 +110,22 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
// Skip freq_cis_real & freq_cis_imag
int head_size = p->dim / p->n_heads;
fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
// Check we didn't forget to read anything
auto curr = ftell(f);
fseek(f, 0, SEEK_END);
auto end = ftell(f);
if (curr != end) {
printf("Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", curr, end);
return 1;
}
return 0;
}
@@ -115,6 +141,7 @@ void free_weights(TransformerWeights* w) {
delete w->w2;
delete w->w3;
delete w->rms_final_weight;
if (w->wcls) delete w->wcls;
}
void print_sample_weights(TransformerWeights *w){
@@ -131,6 +158,7 @@ void print_sample_weights(TransformerWeights *w){
printf("%f\n", w->w2[0]);
printf("%f\n", w->w3[0]);
printf("%f\n", w->rms_att_weight[0]);
if (w->wcls) printf("%f\n", w->wcls[0]);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -509,26 +537,28 @@ bool is_ggml_file(const char *filename) {
}
void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
// 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
#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);
llama_file file(filename, "rb");
const int n_vocab = config->vocab_size;
@@ -538,6 +568,12 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
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;
@@ -589,83 +625,80 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
}
#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);
// }
// 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
stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : 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() {
struct train_params params;
params.fn_vocab_model = "models/ggml-vocab.bin";
params.fn_vocab_model = "tokenizer.bin";
params.fn_llama2c_output_model = "ak_llama_model.bin";
params.fn_train_data = "shakespeare.txt";
params.fn_checkpoint_in = "checkpoint.bin";
@@ -718,7 +751,7 @@ 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 ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
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, " --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");
@@ -791,9 +824,12 @@ int main(int argc, char ** argv) {
if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
// read in the config header
if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
auto shared_weights = config.vocab_size > 0;
config.vocab_size = abs(config.vocab_size);
// read in the Transformer weights
malloc_weights(&weights, &config);
if(checkpoint_init_weights(&weights, &config, file)) { return 1; }
malloc_weights(&weights, &config, shared_weights);
if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
fclose(file);
}
examples/embd-input/embd_input.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import ctypes
from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
import numpy as np
examples/embd-input/llava.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
examples/embd-input/minigpt4.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
examples/embd-input/panda_gpt.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
examples/jeopardy/graph.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import matplotlib.pyplot as plt
import os
import csv
examples/jeopardy/jeopardy.sh
Regular → Executable
View File
examples/json-schema-to-grammar.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import argparse
import json
import re
examples/main/main.cpp
+13 -6
View File
@@ -43,7 +43,7 @@ static bool is_interacting = false;
void sigint_handler(int signo) {
if (signo == SIGINT) {
if (!is_interacting) {
is_interacting=true;
is_interacting = true;
} else {
console::cleanup();
printf("\n");
@@ -189,23 +189,30 @@ int main(int argc, char ** argv) {
}
}
const bool is_spm = 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, true);
embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
} else {
embd_inp = session_tokens;
}
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(ctx));
}
// Tokenize negative prompt
std::vector<llama_token> guidance_inp;
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, true);
guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
}
@@ -252,8 +259,8 @@ int main(int argc, char ** argv) {
}
// prefix & suffix for instruct mode
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm);
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
if (params.instruct) {
examples/make-ggml.py
Regular → Executable
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
"""
This script converts Hugging Face llama models to GGML and quantizes them.
examples/perplexity/perplexity.cpp
+143 -8
View File
@@ -27,12 +27,136 @@ std::vector<float> softmax(const std::vector<float>& logits) {
return probs;
}
void perplexity(llama_context * ctx, const gpt_params & params) {
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`
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
auto tokens = ::llama_tokenize(ctx, params.prompt, true);
if (params.ppl_stride <= 0) {
fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
return;
}
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = is_spm;
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
const int calc_chunk = params.n_ctx;
fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
if (int(tokens.size()) <= calc_chunk) {
fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
tokens.size(), params.n_ctx, params.ppl_stride);
return;
}
const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1) / params.ppl_stride;
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_vocab = llama_n_vocab(ctx);
const int n_batch = params.n_batch;
int count = 0;
double nll = 0.0;
fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
for (int i = 0; i < n_chunk; ++i) {
const int start = i * params.ppl_stride;
const int end = start + calc_chunk;
const int num_batches = (calc_chunk + n_batch - 1) / n_batch;
//fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
std::vector<float> logits;
const auto t_start = std::chrono::high_resolution_clock::now();
for (int j = 0; j < num_batches; ++j) {
const int batch_start = start + j * n_batch;
const int batch_size = std::min(end - batch_start, n_batch);
//fprintf(stderr, " Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
//fprintf(stderr, "%s : failed to eval\n", __func__);
return;
}
// save original token and restore it after eval
const auto token_org = tokens[batch_start];
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[batch_start] = llama_token_bos(ctx);
}
const auto batch_logits = llama_get_logits(ctx);
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
if (j == 0) {
tokens[batch_start] = token_org;
}
}
const auto t_end = std::chrono::high_resolution_clock::now();
if (i == 0) {
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total * n_chunk);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
}
//fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
for (int j = params.n_ctx - params.ppl_stride - 1; 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 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.ppl_stride, std::exp(nll / count));
}
fflush(stdout);
}
printf("\n");
}
void perplexity(llama_context * ctx, const gpt_params & params) {
if (params.ppl_stride > 0) {
perplexity_v2(ctx, params);
return;
}
// 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`
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = is_spm;
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
const int n_chunk_max = tokens.size() / params.n_ctx;
@@ -63,7 +187,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
const auto token_org = tokens[batch_start];
// add BOS token for the first batch of each chunk
if (j == 0) {
if (add_bos && j == 0) {
tokens[batch_start] = llama_token_bos(ctx);
}
@@ -116,7 +240,11 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
++count;
}
// perplexity is e^(average negative log-likelihood)
printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
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));
}
fflush(stdout);
}
printf("\n");
@@ -177,8 +305,10 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
size_t hs_task_count = prompt_lines.size()/6;
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;
// This is needed as usual for LLaMA models
bool prepend_bos = true;
const bool add_bos = is_spm;
// Number of tasks to use when computing the score
if ( params.hellaswag_tasks < hs_task_count ) {
@@ -234,14 +364,13 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
std::vector<float> tok_logits(n_vocab);
for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
// Tokenize the context to count tokens
std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
size_t context_size = context_embd.size();
// 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], prepend_bos);
auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
auto query_size = query_embd.size();
//printf("First query: %d\n",(int)query_size);
@@ -369,6 +498,12 @@ int main(int argc, char ** argv) {
params.perplexity = true;
params.n_batch = std::min(params.n_batch, params.n_ctx);
if (params.ppl_stride > 0) {
fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
params.n_ctx, params.n_ctx + params.ppl_stride/2);
params.n_ctx += params.ppl_stride/2;
}
if (params.n_ctx > 2048) {
fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
"expect poor results\n", __func__, params.n_ctx);
examples/quantize/quantize.cpp
+14 -14
View File
@@ -12,25 +12,25 @@ struct quant_option {
};
static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 3.50G, +0.2499 ppl @ 7B", },
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1846 ppl @ 7B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.30G, +0.0796 ppl @ 7B", },
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0415 ppl @ 7B", },
{ "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 3.56G, +0.2166 ppl @ LLaMA-v1-7B", },
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
#ifdef GGML_USE_K_QUANTS
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.67G, +0.8698 ppl @ 7B", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
{ "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5505 ppl @ 7B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.06G, +0.2437 ppl @ 7B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1803 ppl @ 7B", },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
{ "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
{ "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.56G, +0.1149 ppl @ 7B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0535 ppl @ 7B", },
{ "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
{ "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
{ "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0353 ppl @ 7B", },
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0142 ppl @ 7B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, +0.0044 ppl @ 7B", },
{ "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
#endif
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ 7B", },
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
};
examples/reason-act.sh
-1
View File
@@ -1,4 +1,3 @@
#!/bin/bash
cd `dirname $0`
examples/server-llama2-13B.sh
Regular → Executable
View File
examples/server/api_like_OAI.py
+1
View File
@@ -1,3 +1,4 @@
#!/usr/bin/env python3
import argparse
from flask import Flask, jsonify, request, Response
import urllib.parse
examples/server/chat-llama2.sh
Regular → Executable
View File
examples/server/chat.sh
Regular → Executable
View File