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

add multi adaptor hosting

Browse source
This commit is contained in:
zhhan 2024-07-08 16:05:27 -07:00
parent 5f2d4e60e2
commit 0ab112abdb
8 changed files with 1150 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

30
common/common.cpp
View file

@ -654,6 +654,18 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.hf_file = argv[i];
return true;
}
if (arg == "-mpa" || arg == "--model-path-alias") {
CHECK_ARG
std::string model_derived_alias = argv[i];
size_t equals_pos = model_derived_alias.find('=');
if (equals_pos != std::string::npos) {
std::string alias = model_derived_alias.substr(0, equals_pos);
std::string model_path = model_derived_alias.substr(equals_pos + 1);
params.derived_model_paths.emplace_back(alias, model_path);
}
return true;
}
if (arg == "--lora") {
CHECK_ARG
params.lora_adapter.emplace_back(argv[i], 1.0f);
@ -2045,6 +2057,24 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
}
}
std::map<std::string, llama_model*> derived_models;
for (unsigned int i = 0; i < params.derived_model_paths.size(); ++i) {
const auto & derived_model_path = params.derived_model_paths[i];
const std::string & derived_model_name = std::get<0>(derived_model_path);
const std::string & derived_model_file = std::get<1>(derived_model_path);
llama_model * derived_model_ptr = nullptr;
derived_model_ptr = llama_load_model_from_file(derived_model_file.c_str(), mparams);
if (derived_model_ptr == NULL) {
fprintf(stderr, "%s: error: failed to load derived model '%s'\n", __func__, derived_model_file.c_str());
}
derived_models[derived_model_name] = derived_model_ptr;
}
llama_set_derived_models(lctx, derived_models);
for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
float lora_scale = std::get<1>(params.lora_adapter[i]);

3
common/common.h
View file

@ -124,6 +124,9 @@ struct gpt_params {
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
// multiple derived models paths map
std::vector<std::tuple<std::string, std::string>> derived_model_paths; // derived model paths
// TODO: avoid tuple, use struct
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter

1
examples/CMakeLists.txt
View file

@ -33,6 +33,7 @@ else()
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(main)
add_subdirectory(multi-adaptation)
add_subdirectory(parallel)
add_subdirectory(passkey)
add_subdirectory(perplexity)

5
examples/multi-adaptation/CMakeLists.txt Normal file
View file

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

40
examples/multi-adaptation/README.md Normal file
View file

@ -0,0 +1,40 @@
Server multi adaptations for different scenario.
## Goal
Service multi scenarios on memory constrained devices. The offline models are in the same folder. Use the -mpa parameter to pass the alias and model path. Split the gguf model as below:
## Foundation model
The foundation model contains all the weights parameters used by the runtime. It play as shared split and will be referenced by other gguf models.
model-adaptor-taskA.gguf + model-foundation.gguf
model-adaptor-taskB.gguf + model-foundation.gguf
model-adaptor-base.gguf + model-foundation.gguf
## Model adaptation
Contains partial collections of the model weights that are overlaid onto the foundation model. These adaptation weights can be load dynamically and swapped out based on the usage.
## Example
Use the gguf splits in this model repo: https://huggingface.co/zhhan/Phi-3-mini-4k-instruct_multi-adaptor_gguf
Configuration to run multi-adaptation in visual studio:
{
"type": "default",
"project": "CMakeLists.txt",
"projectTarget": "llama_multi-adaptation.exe (bin\\llama_multi-adaptation.exe)",
"name": "llama_multi-adaptation.exe (bin\\llama_multi-adaptation.exe)",
"args": [
"-ngl 32",
"-m models\phi3_adaptors\\Phi-3-mini-4k-instruct-ft-q4_att-adaptor-base.gguf",
"-mpa codewriter=models\\phi3_adaptors\\Phi-3-mini-4k-instruct-ft-q4_att-adaptor-code_writer.gguf",
"-mpa summarize=models\\phi3_adaptors\\Phi-3-mini-4k-instruct-ft-q4_att-adaptor-summarization.gguf",
"-p \u0022\u003C|user|\u003EHow to explain Internet for a medieval knight?\u003C|end|\u003E\u003C|assistant|\u003E\u0022",
"--color",
"-c 4096",
"--temp 0.7",
"--repeat_penalty 1.1",
"-n 256"
]
}

958
examples/multi-adaptation/multi-adaptation.cpp Normal file
View file

@ -0,0 +1,958 @@
#include "common.h"
#include "console.h"
#include "llama.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <sstream>
#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
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <signal.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static llama_context ** g_ctx;
static llama_model ** g_model;
static gpt_params * g_params;
static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static bool file_exists(const std::string & path) {
std::ifstream f(path.c_str());
return f.good();
}
static bool file_is_empty(const std::string & path) {
std::ifstream f;
f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
return f.tellg() == 0;
}
static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens
) {
if (params.logdir.empty()) {
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Generation Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_dump_timing_info_yaml(logfile, ctx);
fclose(logfile);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
if (!is_interacting && g_params->interactive) {
is_interacting = true;
} else {
console::cleanup();
printf("\n");
llama_print_timings(*g_ctx);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
_exit(130);
}
}
}
#endif
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
llama_chat_msg new_msg{role, content};
auto formatted = llama_chat_format_single(
model, g_params->chat_template, chat_msgs, new_msg, role == "user");
chat_msgs.push_back({role, content});
return formatted;
}
int main(int argc, char ** argv) {
gpt_params params;
g_params = &params;
if (!gpt_params_parse(argc, argv, params)) {
gpt_params_print_usage(argc, argv, params);
return 1;
}
llama_sampling_params & sparams = params.sparams;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("main", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
// TODO: Dump params ?
//LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
// save choice to use color for later
// (note for later: this is a slightly awkward choice)
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
LOG_TEE("%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
if (params.seed == LLAMA_DEFAULT_SEED) {
params.seed = time(NULL);
}
LOG_TEE("%s: seed = %u\n", __func__, params.seed);
std::mt19937 rng(params.seed);
LOG("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
llama_context * ctx_guidance = NULL;
std::vector<llama_chat_msg> chat_msgs;
g_model = &model;
g_ctx = &ctx;
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (sparams.cfg_scale > 1.f) {
struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
ctx_guidance = llama_new_context_with_model(model, lparams);
}
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
}
{
LOG_TEE("\n");
llama_print_derived_models(ctx);
}
llama_switch_derived_model(ctx, "summarize");
std::string path_session = params.path_prompt_cache;
std::vector<llama_token> session_tokens;
if (!path_session.empty()) {
LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
if (!file_exists(path_session)) {
LOG_TEE("%s: session file does not exist, will create.\n", __func__);
} else if (file_is_empty(path_session)) {
LOG_TEE("%s: The session file is empty. A new session will be initialized.\n", __func__);
} else {
// The file exists and is not empty
session_tokens.resize(n_ctx);
size_t n_token_count_out = 0;
if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
return 1;
}
session_tokens.resize(n_token_count_out);
LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
}
}
const bool add_bos = llama_should_add_bos_token(model);
GGML_ASSERT(llama_add_eos_token(model) != 1);
LOG("add_bos: %d\n", add_bos);
std::vector<llama_token> embd_inp;
{
auto prompt = (params.conversation && params.enable_chat_template)
? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
: params.prompt;
if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n");
embd_inp = ::llama_tokenize(ctx, prompt, true, true);
} else {
LOG("use session tokens\n");
embd_inp = session_tokens;
}
LOG("prompt: \"%s\"\n", log_tostr(prompt));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
}
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(model));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
}
// Tokenize negative prompt
std::vector<llama_token> guidance_inp;
int guidance_offset = 0;
int original_prompt_len = 0;
if (ctx_guidance) {
LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true, true);
LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true, true);
LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
LOG("guidance_offset: %s", log_tostr(guidance_offset));
}
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
return 1;
}
// debug message about similarity of saved session, if applicable
size_t n_matching_session_tokens = 0;
if (!session_tokens.empty()) {
for (llama_token id : session_tokens) {
if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
break;
}
n_matching_session_tokens++;
}
if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
LOG_TEE("%s: using full prompt from session file\n", __func__);
} else if (n_matching_session_tokens >= embd_inp.size()) {
LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
} else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
__func__, n_matching_session_tokens, embd_inp.size());
} else {
LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
__func__, n_matching_session_tokens, embd_inp.size());
}
// remove any "future" tokens that we might have inherited from the previous session
llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
}
LOGLN(
"recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
// if we will use the cache for the full prompt without reaching the end of the cache, force
// reevaluation of the last token to recalculate the cached logits
if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
session_tokens.resize(embd_inp.size() - 1);
}
// number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
params.n_keep = (int)embd_inp.size();
} else {
params.n_keep += add_bos; // always keep the BOS token
}
if (params.conversation) {
params.interactive_first = true;
}
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
params.interactive = true;
}
if (params.verbose_prompt) {
LOG_TEE("\n");
LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (ctx_guidance) {
LOG_TEE("\n");
LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
for (int i = 0; i < (int) guidance_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
}
}
if (params.n_keep > add_bos) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
}
LOG_TEE("\n");
}
// ctrl+C handling
{
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
struct sigaction sigint_action;
sigint_action.sa_handler = sigint_handler;
sigemptyset (&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined (_WIN32)
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
}
if (params.interactive) {
LOG_TEE("%s: interactive mode on.\n", __func__);
if (!params.antiprompt.empty()) {
for (const auto & antiprompt : params.antiprompt) {
LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
}
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
}
LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
// group-attention state
// number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
int ga_i = 0;
const int ga_n = params.grp_attn_n;
const int ga_w = params.grp_attn_w;
if (ga_n != 1) {
GGML_ASSERT(ga_n > 0 && "grp_attn_n must be positive"); // NOLINT
GGML_ASSERT(ga_w % ga_n == 0 && "grp_attn_w must be a multiple of grp_attn_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
}
LOG_TEE("\n\n");
if (params.interactive) {
const char * control_message;
if (params.multiline_input) {
control_message = " - To return control to the AI, end your input with '\\'.\n"
" - To return control without starting a new line, end your input with '/'.\n";
} else {
control_message = " - Press Return to return control to the AI.\n"
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
LOG_TEE("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
is_interacting = params.interactive_first;
}
bool is_antiprompt = false;
bool input_echo = true;
bool display = true;
bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
int n_past = 0;
int n_remain = params.n_predict;
int n_consumed = 0;
int n_session_consumed = 0;
int n_past_guidance = 0;
std::vector<int> input_tokens; g_input_tokens = &input_tokens;
std::vector<int> output_tokens; g_output_tokens = &output_tokens;
std::ostringstream output_ss; g_output_ss = &output_ss;
std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
// the first thing we will do is to output the prompt, so set color accordingly
console::set_display(console::prompt);
display = params.display_prompt;
std::vector<llama_token> embd;
std::vector<llama_token> embd_guidance;
// tokenized antiprompts
std::vector<std::vector<llama_token>> antiprompt_ids;
antiprompt_ids.reserve(params.antiprompt.size());
for (const std::string & antiprompt : params.antiprompt) {
antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
}
struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
if (!ctx_sampling) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
// predict
if (!embd.empty()) {
// Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
// --prompt or --file which uses the same value.
int max_embd_size = n_ctx - 4;
// Ensure the input doesn't exceed the context size by truncating embd if necessary.
if ((int) embd.size() > max_embd_size) {
const int skipped_tokens = (int) embd.size() - max_embd_size;
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
if (ga_n == 1) {
// infinite text generation via context shifting
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) >= n_ctx) {
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
if (ctx_guidance) {
n_past_guidance -= n_discard;
}
LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG("clear session path\n");
path_session.clear();
}
} else {
// context extension via Self-Extend
while (n_past >= ga_i + ga_w) {
const int ib = (ga_n*ga_i)/ga_w;
const int bd = (ga_w/ga_n)*(ga_n - 1);
const int dd = (ga_w/ga_n) - ib*bd - ga_w;
LOG("\n");
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
LOG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd);
llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);
llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd);
n_past -= bd;
ga_i += ga_w/ga_n;
LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
}
}
// try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
if (n_session_consumed < (int) session_tokens.size()) {
size_t i = 0;
for ( ; i < embd.size(); i++) {
if (embd[i] != session_tokens[n_session_consumed]) {
session_tokens.resize(n_session_consumed);
break;
}
n_past++;
n_session_consumed++;
if (n_session_consumed >= (int) session_tokens.size()) {
++i;
break;
}
}
if (i > 0) {
embd.erase(embd.begin(), embd.begin() + i);
}
}
// evaluate tokens in batches
// embd is typically prepared beforehand to fit within a batch, but not always
if (ctx_guidance) {
int input_size = 0;
llama_token * input_buf = NULL;
if (n_past_guidance < (int) guidance_inp.size()) {
// Guidance context should have the same data with these modifications:
//
// * Replace the initial prompt
// * Shift everything by guidance_offset
embd_guidance = guidance_inp;
if (embd.begin() + original_prompt_len < embd.end()) {
embd_guidance.insert(
embd_guidance.end(),
embd.begin() + original_prompt_len,
embd.end()
);
}
input_buf = embd_guidance.data();
input_size = embd_guidance.size();
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
} else {
input_buf = embd.data();
input_size = embd.size();
}
for (int i = 0; i < input_size; i += params.n_batch) {
int n_eval = std::min(input_size - i, params.n_batch);
if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
return 1;
}
n_past_guidance += n_eval;
}
}
for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
int n_eval = (int) embd.size() - i;
if (n_eval > params.n_batch) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
// Display total tokens alongside total time
if (params.n_print > 0 && n_past % params.n_print == 0) {
LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
}
}
if (!embd.empty() && !path_session.empty()) {
session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
n_session_consumed = session_tokens.size();
}
}
embd.clear();
embd_guidance.clear();
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
// optionally save the session on first sample (for faster prompt loading next time)
if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
need_to_save_session = false;
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
LOG("saved session to %s\n", path_session.c_str());
}
const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
embd.push_back(id);
// echo this to console
input_echo = true;
// decrement remaining sampling budget
--n_remain;
LOG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
while ((int) embd_inp.size() > n_consumed) {
embd.push_back(embd_inp[n_consumed]);
// push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
break;
}
}
}
// display text
if (input_echo && display) {
for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id, params.special);
// Console/Stream Output
fprintf(stdout, "%s", token_str.c_str());
// Record Displayed Tokens To Log
// Note: Generated tokens are created one by one hence this check
if (embd.size() > 1) {
// Incoming Requested Tokens
input_tokens.push_back(id);
} else {
// Outgoing Generated Tokens
output_tokens.push_back(id);
output_ss << token_str;
}
fflush(stdout);
}
}
// reset color to default if there is no pending user input
if (input_echo && (int) embd_inp.size() == n_consumed) {
console::set_display(console::reset);
display = true;
}
// if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) {
// check for reverse prompt in the last n_prev tokens
if (!params.antiprompt.empty()) {
const int n_prev = 32;
const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
is_antiprompt = false;
// Check if each of the reverse prompts appears at the end of the output.
// If we're not running interactively, the reverse prompt might be tokenized with some following characters
// so we'll compensate for that by widening the search window a bit.
for (std::string & antiprompt : params.antiprompt) {
size_t extra_padding = params.interactive ? 0 : 2;
size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)
? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
: 0;
if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
}
}
// check for reverse prompt using special tokens
llama_token last_token = llama_sampling_last(ctx_sampling);
for (std::vector<llama_token> ids : antiprompt_ids) {
if (ids.size() == 1 && last_token == ids[0]) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
}
}
if (is_antiprompt) {
LOG("found antiprompt: %s\n", last_output.c_str());
}
}
// deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
LOG("found an EOG token\n");
if (params.interactive) {
if (!params.antiprompt.empty()) {
// tokenize and inject first reverse prompt
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
is_antiprompt = true;
}
if (params.enable_chat_template) {
chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
}
is_interacting = true;
printf("\n");
}
}
// if current token is not EOG, we add it to current assistant message
if (params.conversation) {
auto id = llama_sampling_last(ctx_sampling);
assistant_ss << llama_token_to_piece(ctx, id, false);
}
if (n_past > 0 && is_interacting) {
LOG("waiting for user input\n");
if (params.conversation) {
printf("\n> ");
}
if (params.input_prefix_bos) {
LOG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
}
std::string buffer;
if (!params.input_prefix.empty() && !params.conversation) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
printf("%s", params.input_prefix.c_str());
}
// color user input only
console::set_display(console::user_input);
display = params.display_prompt;
std::string line;
bool another_line = true;
do {
another_line = console::readline(line, params.multiline_input);
buffer += line;
} while (another_line);
// done taking input, reset color
console::set_display(console::reset);
display = true;
// Add tokens to embd only if the input buffer is non-empty
// Entering a empty line lets the user pass control back
if (buffer.length() > 1) {
// append input suffix if any
if (!params.input_suffix.empty() && !params.conversation) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
printf("%s", params.input_suffix.c_str());
}
LOG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
if (params.escape) {
string_process_escapes(buffer);
}
bool format_chat = params.conversation && params.enable_chat_template;
std::string user_inp = format_chat
? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
: std::move(buffer);
// TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
const auto line_inp = ::llama_tokenize(ctx, user_inp, false, format_chat);
const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i];
output_tokens.push_back(token);
output_ss << llama_token_to_piece(ctx, token);
}
// reset assistant message
assistant_ss.str("");
n_remain -= line_inp.size();
LOG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
}
if (n_past > 0) {
if (is_interacting) {
llama_sampling_reset(ctx_sampling);
}
is_interacting = false;
}
}
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
LOG_TEE(" [end of text]\n");
break;
}
// In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
// 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;
}
}
if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
}
llama_print_timings(ctx);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
if (ctx_guidance) { llama_free(ctx_guidance); }
llama_free(ctx);
llama_free_model(model);
llama_sampling_free(ctx_sampling);
llama_backend_free();
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
return 0;
}

19
include/llama.h
View file

@ -8,6 +8,8 @@
#include <stdint.h>
#include <stdio.h>
#include <stdbool.h>
#include <map>
#include <string>
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
@ -425,6 +427,18 @@ extern "C" {
struct llama_model * model,
struct llama_context_params params);
LLAMA_API void llama_print_derived_models(struct llama_context* ctx);
LLAMA_API void llama_set_derived_models(
struct llama_context * ctx,
std::map<std::string, struct llama_model *> derived_models);
static const char* BASE_MODEL = "base";
LLAMA_API bool llama_switch_derived_model(
struct llama_context* ctx,
std::string derived_model_name);
// Frees all allocated memory
LLAMA_API void llama_free(struct llama_context * ctx);
@ -1087,6 +1101,11 @@ extern "C" {
// Returns the split_prefix length.
LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
LLAMA_API int llama_foundation_split_path(char* split_path, size_t maxlen, const char* path_prefix);
LLAMA_API int llama_foundation_prefix(char* split_path, size_t maxlen, const char* path_prefix);
// Performance information
LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);

101
src/llama.cpp
View file

@ -2248,6 +2248,8 @@ struct llama_cparams {
bool offload_kqv;
bool flash_attn;
std::string derived_model_name = BASE_MODEL;
enum llama_pooling_type pooling_type;
ggml_backend_sched_eval_callback cb_eval;
@ -2621,6 +2623,9 @@ struct llama_context {
const llama_model & model;
// derived models
std::map<std::string, llama_model *> derived_models;
// key + value cache for the self attention
struct llama_kv_cache kv_self;
@ -3543,7 +3548,14 @@ struct llama_model_loader {
}
char split_prefix[PATH_MAX] = {0};
if (!llama_split_prefix(split_prefix, sizeof(split_prefix), fname.c_str(), idx, n_split)) {
char foundation_prefix[PATH_MAX] = { 0 };
// Two types of split files are supported:
// prefix is abc, postfix is 00001-of-00002, 00002-of-00002
// abc-00001-of-00002.gguf, abc-00002-of-00002.gguf
// prefix is abc, postfix is foundation, adaptor-task-x, adaptor-task-y
// abc-foundation.gguf, abc-adaptor-task-x.gguf, abc-adaptor-task-y.gguf
if (!llama_split_prefix(split_prefix, sizeof(split_prefix), fname.c_str(), idx, n_split)
&& !llama_foundation_prefix(foundation_prefix, sizeof(foundation_prefix), fname.c_str())) {
throw std::runtime_error(format("invalid split file: %s", fname.c_str()));
}
@ -3555,6 +3567,14 @@ struct llama_model_loader {
for (idx = 1; idx < n_split; idx++) {
llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
// if split path not exist
struct stat model_file_info;
std::string str_split_path(split_path);
auto file_exists = (stat(str_split_path.c_str(), &model_file_info) == 0);
if (!file_exists) {
llama_foundation_split_path(split_path, sizeof(split_path), foundation_prefix);
}
struct gguf_init_params split_params = {
/*.no_alloc = */ true,
/*.ctx = */ &ctx,
@ -7804,10 +7824,11 @@ struct llm_build_context {
// TODO: consider making the entire interface noexcept
llm_build_context(
llama_context & lctx,
const llama_model & model,
const llama_batch & batch,
const llm_build_cb & cb,
bool worst_case) :
model (lctx.model),
model (model),
lctx (lctx),
hparams (model.hparams),
cparams (lctx.cparams),
@ -12525,7 +12546,7 @@ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const
llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, lctx.model, dummy, cb, false);
llm.init();
@ -12542,7 +12563,7 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, lctx.model, dummy, cb, false);
llm.init();
@ -12559,7 +12580,7 @@ static struct ggml_cgraph * llama_build_graph_s_copy(llama_context & lctx) {
llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, lctx.model, dummy, cb, false);
llm.init();
@ -12574,7 +12595,19 @@ static struct ggml_cgraph * llama_build_graph(
llama_context & lctx,
const llama_batch & batch,
bool worst_case) {
const auto & model = lctx.model;
const auto & foundation_model = lctx.model;
const llama_model* model_ptr = nullptr;
const auto it = lctx.derived_models.find(lctx.cparams.derived_model_name);
if (it != lctx.derived_models.end()) {
const auto& model_derived = *(it->second);
model_ptr = &model_derived;
}
else {
model_ptr = &foundation_model;
}
const llama_model & model = *model_ptr;
// this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) {
@ -12609,7 +12642,7 @@ static struct ggml_cgraph * llama_build_graph(
struct ggml_cgraph * result = NULL;
struct llm_build_context llm(lctx, batch, cb, worst_case);
struct llm_build_context llm(lctx, lctx.model, batch, cb, worst_case);
llm.init();
@ -17955,6 +17988,17 @@ struct llama_context * llama_new_context_with_model(
return ctx;
}
struct llama_context * llama_new_context_with_derived_models(
struct llama_model * model,
struct llama_context_params params,
const std::map<std::string, llama_model*> derived_models) {
llama_context * ctx = llama_new_context_with_model(model, params);
if (ctx) {
ctx->derived_models = derived_models;
}
return ctx;
}
void llama_free(struct llama_context * ctx) {
delete ctx;
}
@ -17963,6 +18007,26 @@ const llama_model * llama_get_model(const struct llama_context * ctx) {
return &ctx->model;
}
void llama_print_derived_models(struct llama_context * ctx) {
for (const auto & it : ctx->derived_models) {
LLAMA_LOG_INFO("%s: %s\n", __func__, it.first.c_str());
}
}
void llama_set_derived_models(struct llama_context * ctx, const std::map<std::string, llama_model*> derived_models) {
ctx->derived_models = derived_models;
}
bool llama_switch_derived_model(struct llama_context* ctx, const std::string derived_model_name) {
llama_synchronize(ctx);
auto& cparams = ctx->cparams;
cparams.derived_model_name = (ctx->derived_models.find(derived_model_name) == ctx->derived_models.end()) ? BASE_MODEL : derived_model_name;
LLAMA_LOG_INFO("%s: %s\n", __func__, cparams.derived_model_name);
return true;
}
uint32_t llama_n_ctx(const struct llama_context * ctx) {
return ctx->cparams.n_ctx;
}
@ -20111,6 +20175,29 @@ LLAMA_API int32_t llama_chat_apply_template(
return res;
}
LLAMA_API int llama_foundation_split_path(char* split_path, size_t maxlen, const char* path_prefix) {
static const char* const SHARED_SPLIT_PATH_FORMAT = "%s-foundation.gguf";
if (snprintf(split_path, maxlen, SHARED_SPLIT_PATH_FORMAT, path_prefix)) {
return strlen(split_path);
}
return 0;
}
int llama_foundation_prefix(char* dest, size_t maxlen, const char* split_path) {
const char* keyword = "-adaptor-";
const char* pos = strstr(split_path, keyword);
if (pos != NULL) {
size_t size_prefix = pos - split_path;
snprintf(dest, std::min((size_t)size_prefix + 1, maxlen), "%s", split_path);
// strncpy(dest, split_path, size_prefix);
// dest[size_prefix] = '\0';
return size_prefix;
}
return 0;
}
LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {