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

Merge remote-tracking branch 'upstream/master' into nomic-vulkan-redo

Browse source
This commit is contained in:
Jared Van Bortel 2023-11-23 13:05:04 -05:00
commit 9ae88baf38
50 changed files with 1949 additions and 1693 deletions
Download patch file Download diff file Expand all files Collapse all files

1
examples/CMakeLists.txt
View file

@ -24,6 +24,7 @@ else()
add_subdirectory(llama-bench)
add_subdirectory(llava)
add_subdirectory(main)
add_subdirectory(tokenize)
add_subdirectory(parallel)
add_subdirectory(perplexity)
add_subdirectory(quantize)

2
examples/finetune/README.md
View file

@ -21,7 +21,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
```
Finetune output files will be saved every N iterations (config with `--save-every N`).
**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
So in above example after 10 iterations these files will be written:
- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf

2
examples/finetune/convert-finetune-checkpoint-to-gguf.py
View file

@ -3,9 +3,7 @@
import argparse
import gguf
import os
import struct
import sys
import numpy as np
from pathlib import Path

35
examples/finetune/finetune.cpp
View file

@ -548,35 +548,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
randomize_tensor_normal(lora->tok_embeddings_a, rnd);
randomize_tensor_normal(lora->tok_embeddings_b, rnd);
ggml_set_zero(lora->tok_embeddings_b);
randomize_tensor_normal(lora->norm_a, rnd);
randomize_tensor_normal(lora->norm_b, rnd);
ggml_set_zero(lora->norm_b);
randomize_tensor_normal(lora->output_a, rnd);
randomize_tensor_normal(lora->output_b, rnd);
ggml_set_zero(lora->output_b);
for (uint32_t i = 0; i < n_layer; ++i) {
auto & layer = lora->layers[i];
randomize_tensor_normal(layer.attention_norm_a, rnd);
randomize_tensor_normal(layer.attention_norm_b, rnd);
ggml_set_zero(layer.attention_norm_b);
randomize_tensor_normal(layer.wq_a, rnd);
randomize_tensor_normal(layer.wq_b, rnd);
ggml_set_zero(layer.wq_b);
randomize_tensor_normal(layer.wk_a, rnd);
randomize_tensor_normal(layer.wk_b, rnd);
ggml_set_zero(layer.wk_b);
randomize_tensor_normal(layer.wv_a, rnd);
randomize_tensor_normal(layer.wv_b, rnd);
ggml_set_zero(layer.wv_b);
randomize_tensor_normal(layer.wo_a, rnd);
randomize_tensor_normal(layer.wo_b, rnd);
ggml_set_zero(layer.wo_b);
randomize_tensor_normal(layer.ffn_norm_a, rnd);
randomize_tensor_normal(layer.ffn_norm_b, rnd);
ggml_set_zero(layer.ffn_norm_b);
randomize_tensor_normal(layer.w1_a, rnd);
randomize_tensor_normal(layer.w1_b, rnd);
ggml_set_zero(layer.w1_b);
randomize_tensor_normal(layer.w2_a, rnd);
randomize_tensor_normal(layer.w2_b, rnd);
ggml_set_zero(layer.w2_b);
randomize_tensor_normal(layer.w3_a, rnd);
randomize_tensor_normal(layer.w3_b, rnd);
ggml_set_zero(layer.w3_b);
}
free_random_normal_distribution(rnd);
@ -1460,17 +1460,6 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
}
params->n_rank_w3 = std::stoi(argv[i]);
params->custom_n_rank_w3 = true;
} else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
if (++i >= argc) {
invalid_param = true;
break;
}
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
params->common.n_gpu_layers = std::stoi(argv[i]);
#else
fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
#endif
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
train_print_usage(argc, argv, &default_params);

9
examples/infill/infill.cpp
View file

@ -146,6 +146,13 @@ int main(int argc, char ** argv) {
return 0;
}
if (params.chatml) {
printf("\n************\n");
printf("%s: please use the 'main' tool for chatml mode\n", __func__);
printf("************\n\n");
return 0;
}
if (!params.antiprompt.empty()) {
printf("\n************\n");
printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
@ -230,7 +237,7 @@ int main(int argc, char ** argv) {
LOG_TEE("\n");
LOG_TEE("%s\n", get_system_info(params).c_str());
}
const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = llama_should_add_bos_token(model);
LOG("add_bos: %d\n", add_bos);
bool suff_rm_leading_spc = params.escape;

2
examples/llava/clip.cpp
View file

@ -761,7 +761,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
temp->ny = img->ny;
temp->size = img->size;
temp->data = new uint8_t[temp->size]();
*temp->data = *img->data; // copy
memcpy(&temp->data[0], &img->data[0], temp->size); // copy
}
const int nx = temp->nx;

3
examples/llava/llava-cli.cpp
View file

@ -208,9 +208,10 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
int n_past = 0;
const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
// llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, true);
eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos);
llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);

9
examples/llava/llava.cpp
View file

@ -127,7 +127,14 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
fclose(file);
return false;
}
fread(buffer, 1, fileSize, file); // Read the file into the buffer
errno = 0;
size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
if (ferror(file)) {
die_fmt("read error: %s", strerror(errno));
}
if (ret != (size_t) fileSize) {
die("unexpectedly reached end of file");
}
fclose(file); // Close the file
*bytesOut = buffer;

38
examples/main/main.cpp
View file

@ -237,13 +237,16 @@ int main(int argc, char ** argv) {
}
}
const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = llama_should_add_bos_token(model);
LOG("add_bos: %d\n", add_bos);
std::vector<llama_token> embd_inp;
if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
if (params.interactive_first || params.instruct || params.chatml || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n");
if (params.chatml) {
params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>";
}
embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
} else {
LOG("use session tokens\n");
@ -321,7 +324,7 @@ int main(int argc, char ** argv) {
}
// number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) {
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
params.n_keep = (int)embd_inp.size();
}
@ -332,11 +335,23 @@ int main(int argc, char ** argv) {
LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
// chatml prefix & suffix
const auto cml_pfx = ::llama_tokenize(ctx, "\n<|im_start|>user\n", add_bos, true);
const auto cml_sfx = ::llama_tokenize(ctx, "<|im_end|>\n<|im_start|>assistant\n", false, true);
LOG("cml_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_pfx).c_str());
LOG("cml_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_sfx).c_str());
// in instruct mode, we inject a prefix and a suffix to each input by the user
if (params.instruct) {
params.interactive_first = true;
params.antiprompt.push_back("### Instruction:\n\n");
}
// similar for chatml mode
else if (params.chatml) {
params.interactive_first = true;
params.antiprompt.push_back("<|im_start|>user\n");
}
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
@ -713,7 +728,7 @@ int main(int argc, char ** argv) {
is_interacting = true;
printf("\n");
} else if (params.instruct) {
} else if (params.instruct || params.chatml) {
is_interacting = true;
}
}
@ -721,7 +736,7 @@ int main(int argc, char ** argv) {
if (n_past > 0 && is_interacting) {
LOG("waiting for user input\n");
if (params.instruct) {
if (params.instruct || params.chatml) {
printf("\n> ");
}
@ -768,6 +783,12 @@ int main(int argc, char ** argv) {
n_consumed = embd_inp.size();
embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
}
// chatml mode: insert user chat prefix
if (params.chatml && !is_antiprompt) {
LOG("inserting chatml prefix\n");
n_consumed = embd_inp.size();
embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
}
if (params.escape) {
process_escapes(buffer);
}
@ -786,6 +807,11 @@ int main(int argc, char ** argv) {
LOG("inserting instruction suffix\n");
embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
}
// chatml mode: insert assistant chat suffix
if (params.chatml) {
LOG("inserting chatml suffix\n");
embd_inp.insert(embd_inp.end(), cml_sfx.begin(), cml_sfx.end());
}
for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i];
@ -811,7 +837,7 @@ int main(int argc, char ** argv) {
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) {
if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive || params.chatml)) {
LOG_TEE(" [end of text]\n");
break;
}

11
examples/parallel/parallel.cpp
View file

@ -1,5 +1,5 @@
// A basic application simulating a server with multiple clients.
// The clients submite requests to the server and they are processed in parallel.
// The clients submit requests to the server and they are processed in parallel.
#include "common.h"
#include "llama.h"
@ -113,6 +113,8 @@ int main(int argc, char ** argv) {
// insert new requests as soon as the previous one is done
const bool cont_batching = params.cont_batching;
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("parallel", "log"));
LOG_TEE("Log start\n");
@ -172,6 +174,8 @@ int main(int argc, char ** argv) {
int32_t n_total_gen = 0;
int32_t n_cache_miss = 0;
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients);
const auto t_main_start = ggml_time_us();
LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
@ -201,6 +205,11 @@ int main(int argc, char ** argv) {
LOG_TEE("Processing requests ...\n\n");
while (true) {
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
dump_kv_cache_view_seqs(kvc_view, 40);
}
llama_batch_clear(batch);
// decode any currently ongoing sequences

8
examples/perplexity/perplexity.cpp
View file

@ -149,8 +149,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = is_spm;
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
@ -288,8 +287,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
const bool add_bos = is_spm;
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
const int n_ctx = llama_n_ctx(ctx);
auto tim1 = std::chrono::high_resolution_clock::now();
@ -481,7 +479,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
fprintf(stderr, "================================= is_spm = %d\n", is_spm);
// This is needed as usual for LLaMA models
const bool add_bos = is_spm;
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
// Number of tasks to use when computing the score
if ( params.hellaswag_tasks < hs_task_count ) {

4
examples/server/public/completion.js
View file

@ -94,6 +94,10 @@ export async function* llama(prompt, params = {}, config = {}) {
break;
}
}
if (result.error) {
result.error = JSON.parse(result.error);
console.error(`llama.cpp error: ${result.error.content}`);
}
}
}
}

20
examples/server/server.cpp
View file

@ -501,6 +501,7 @@ struct llama_server_context
bool multimodal = false;
bool clean_kv_cache = true;
bool all_slots_are_idle = false;
bool add_bos_token = true;
int32_t id_gen;
int32_t n_ctx; // total context for all clients / slots
@ -573,6 +574,8 @@ struct llama_server_context
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
return true;
}
@ -864,7 +867,7 @@ struct llama_server_context
}
void update_system_prompt() {
system_tokens = ::llama_tokenize(ctx, system_prompt, true);
system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
llama_batch_clear(batch);
@ -1552,7 +1555,7 @@ struct llama_server_context
}
else
{
prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt
prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
}
slot.num_prompt_tokens = prompt_tokens.size();
@ -1629,7 +1632,7 @@ struct llama_server_context
const bool has_images = process_images(slot);
// process the prefix of first image
std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, true) : prompt_tokens;
std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past)
{
llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false);
@ -2365,6 +2368,17 @@ int main(int argc, char **argv)
break;
}
} else {
const std::string str =
"error: " +
result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
if (!sink.write(str.c_str(), str.size()))
{
return false;
}
break;
}
}

17
examples/speculative/speculative.cpp
View file

@ -94,9 +94,22 @@ int main(int argc, char ** argv) {
}
}
// tokenize the prompt
// Tokenize the prompt
const bool add_bos_tgt = llama_should_add_bos_token(model_tgt);
LOG("add_bos tgt: %d\n", add_bos_tgt);
const bool add_bos_dft = llama_should_add_bos_token(model_dft);
LOG("add_bos dft: %d\n", add_bos_dft);
if (add_bos_tgt != add_bos_dft) {
fprintf(stderr, "%s: error: draft model add_bos must match target model to use speculation but ", __func__);
fprintf(stderr, "add_bos_dft = %d while add_bos_tgt = %d\n", add_bos_dft, add_bos_tgt);
return 1;
}
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
inp = ::llama_tokenize(ctx_tgt, params.prompt, add_bos_tgt, true);
const int max_context_size = llama_n_ctx(ctx_tgt);
const int max_tokens_list_size = max_context_size - 4;

5
examples/tokenize/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET tokenize)
add_executable(${TARGET} tokenize.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

44
examples/tokenize/tokenize.cpp Normal file
View file

@ -0,0 +1,44 @@
#include "common.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
int main(int argc, char ** argv) {
if (argc < 3 || argv[1][0] == '-') {
printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]);
return 1;
}
const char * model_path = argv[1];
const char * prompt = argv[2];
const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
llama_backend_init(false);
llama_model_params model_params = llama_model_default_params();
model_params.vocab_only = true;
llama_model * model = llama_load_model_from_file(model_path, model_params);
llama_context_params ctx_params = llama_context_default_params();
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
const bool add_bos = llama_should_add_bos_token(model);
std::vector<llama_token> tokens;
tokens = ::llama_tokenize(model, prompt, add_bos, true);
for (int i = 0; i < (int) tokens.size(); i++) {
if (printing_ids) {
printf("%d\n", tokens[i]);
} else {
printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str());
}
}
return 0;
}