KBaba7/llama.cpp
0
1#include "arg.h"2#include "common.h"3#include "console.h"4#include "log.h"5#include "sampling.h"6#include "llama.h"7#include "chat-template.hpp"8 9#include <cstdio>10#include <cstring>11#include <ctime>12#include <fstream>13#include <iostream>14#include <sstream>15#include <string>16#include <vector>17 18#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))19#include <signal.h>20#include <unistd.h>21#elif defined (_WIN32)22#define WIN32_LEAN_AND_MEAN23#ifndef NOMINMAX24#define NOMINMAX25#endif26#include <windows.h>27#include <signal.h>28#endif29 30#if defined(_MSC_VER)31#pragma warning(disable: 4244 4267) // possible loss of data32#endif33 34static const char * DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant";35 36static llama_context ** g_ctx;37static llama_model ** g_model;38static common_sampler ** g_smpl;39static common_params * g_params;40static std::vector<llama_token> * g_input_tokens;41static std::ostringstream * g_output_ss;42static std::vector<llama_token> * g_output_tokens;43static bool is_interacting = false;44static bool need_insert_eot = false;45 46static void print_usage(int argc, char ** argv) {47 (void) argc;48 49 LOG("\nexample usage:\n");50 LOG("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);51 LOG("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);52 LOG("\n");53}54 55static bool file_exists(const std::string & path) {56 std::ifstream f(path.c_str());57 return f.good();58}59 60static bool file_is_empty(const std::string & path) {61 std::ifstream f;62 f.exceptions(std::ifstream::failbit | std::ifstream::badbit);63 f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);64 return f.tellg() == 0;65}66 67#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)68static void sigint_handler(int signo) {69 if (signo == SIGINT) {70 if (!is_interacting && g_params->interactive) {71 is_interacting = true;72 need_insert_eot = true;73 } else {74 console::cleanup();75 LOG("\n");76 common_perf_print(*g_ctx, *g_smpl);77 78 // make sure all logs are flushed79 LOG("Interrupted by user\n");80 common_log_pause(common_log_main());81 82 _exit(130);83 }84 }85}86#endif87 88int main(int argc, char ** argv) {89 common_params params;90 g_params = ¶ms;91 if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {92 return 1;93 }94 95 common_init();96 97 auto & sparams = params.sampling;98 99 // save choice to use color for later100 // (note for later: this is a slightly awkward choice)101 console::init(params.simple_io, params.use_color);102 atexit([]() { console::cleanup(); });103 104 if (params.logits_all) {105 LOG_ERR("************\n");106 LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);107 LOG_ERR("************\n\n");108 109 return 0;110 }111 112 if (params.embedding) {113 LOG_ERR("************\n");114 LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);115 LOG_ERR("************\n\n");116 117 return 0;118 }119 120 if (params.n_ctx != 0 && params.n_ctx < 8) {121 LOG_WRN("%s: warning: minimum context size is 8, using minimum size.\n", __func__);122 params.n_ctx = 8;123 }124 125 if (params.rope_freq_base != 0.0) {126 LOG_WRN("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);127 }128 129 if (params.rope_freq_scale != 0.0) {130 LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);131 }132 133 LOG_INF("%s: llama backend init\n", __func__);134 135 llama_backend_init();136 llama_numa_init(params.numa);137 138 llama_model * model = nullptr;139 llama_context * ctx = nullptr;140 common_sampler * smpl = nullptr;141 142 g_model = &model;143 g_ctx = &ctx;144 g_smpl = &smpl;145 146 std::vector<common_chat_msg> chat_msgs;147 148 // load the model and apply lora adapter, if any149 LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);150 common_init_result llama_init = common_init_from_params(params);151 152 model = llama_init.model.get();153 ctx = llama_init.context.get();154 155 if (model == NULL) {156 LOG_ERR("%s: error: unable to load model\n", __func__);157 return 1;158 }159 160 const llama_vocab * vocab = llama_model_get_vocab(model);161 auto chat_templates = common_chat_templates_from_model(model, params.chat_template);162 163 LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);164 165 auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));166 auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");167 auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");168 169 struct ggml_threadpool_params tpp_batch =170 ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);171 struct ggml_threadpool_params tpp =172 ggml_threadpool_params_from_cpu_params(params.cpuparams);173 174 set_process_priority(params.cpuparams.priority);175 176 struct ggml_threadpool * threadpool_batch = NULL;177 if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {178 threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);179 if (!threadpool_batch) {180 LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);181 return 1;182 }183 184 // Start the non-batch threadpool in the paused state185 tpp.paused = true;186 }187 188 struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);189 if (!threadpool) {190 LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);191 return 1;192 }193 194 llama_attach_threadpool(ctx, threadpool, threadpool_batch);195 196 const int n_ctx_train = llama_model_n_ctx_train(model);197 const int n_ctx = llama_n_ctx(ctx);198 199 if (n_ctx > n_ctx_train) {200 LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);201 }202 203 // auto enable conversation mode if chat template is available204 const bool has_chat_template = chat_templates.has_explicit_template && chat_templates.template_default;205 if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {206 if (has_chat_template) {207 LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);208 params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;209 } else {210 params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;211 }212 }213 214 // in case user force-activate conversation mode (via -cnv) without proper chat template, we show a warning215 if (params.conversation_mode && !has_chat_template) {216 LOG_WRN("%s: chat template is not available or is not supported. This may cause the model to output suboptimal responses\n", __func__);217 }218 219 // print chat template example in conversation mode220 if (params.conversation_mode) {221 if (params.enable_chat_template) {222 LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(*chat_templates.template_default, params.use_jinja).c_str());223 } else {224 LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);225 }226 }227 228 // print system information229 {230 LOG_INF("\n");231 LOG_INF("%s\n", common_params_get_system_info(params).c_str());232 LOG_INF("\n");233 }234 235 std::string path_session = params.path_prompt_cache;236 std::vector<llama_token> session_tokens;237 238 if (!path_session.empty()) {239 LOG_INF("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());240 if (!file_exists(path_session)) {241 LOG_INF("%s: session file does not exist, will create.\n", __func__);242 } else if (file_is_empty(path_session)) {243 LOG_INF("%s: The session file is empty. A new session will be initialized.\n", __func__);244 } else {245 // The file exists and is not empty246 session_tokens.resize(n_ctx);247 size_t n_token_count_out = 0;248 if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {249 LOG_ERR("%s: failed to load session file '%s'\n", __func__, path_session.c_str());250 return 1;251 }252 session_tokens.resize(n_token_count_out);253 LOG_INF("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());254 }255 }256 257 const bool add_bos = llama_vocab_get_add_bos(vocab) && !params.use_jinja;258 if (!llama_model_has_encoder(model)) {259 GGML_ASSERT(!llama_vocab_get_add_eos(vocab));260 }261 262 LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);263 264 std::vector<llama_token> embd_inp;265 266 auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {267 common_chat_msg new_msg{role, content, {}};268 auto formatted = common_chat_format_single(*chat_templates.template_default, chat_msgs, new_msg, role == "user", g_params->use_jinja);269 chat_msgs.push_back({role, content, {}});270 LOG_DBG("formatted: '%s'\n", formatted.c_str());271 return formatted;272 };273 274 {275 auto prompt = (params.conversation_mode && params.enable_chat_template)276 // format the system prompt in conversation mode (fallback to default if empty)277 ? chat_add_and_format("system", params.prompt.empty() ? DEFAULT_SYSTEM_MESSAGE : params.prompt)278 // otherwise use the prompt as is279 : params.prompt;280 if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {281 LOG_DBG("tokenize the prompt\n");282 embd_inp = common_tokenize(ctx, prompt, true, true);283 } else {284 LOG_DBG("use session tokens\n");285 embd_inp = session_tokens;286 }287 288 LOG_DBG("prompt: \"%s\"\n", prompt.c_str());289 LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());290 }291 292 // Should not run without any tokens293 if (embd_inp.empty()) {294 if (add_bos) {295 embd_inp.push_back(llama_vocab_bos(vocab));296 LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());297 } else {298 LOG_ERR("input is empty\n");299 return -1;300 }301 }302 303 // Tokenize negative prompt304 if ((int) embd_inp.size() > n_ctx - 4) {305 LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);306 return 1;307 }308 309 // debug message about similarity of saved session, if applicable310 size_t n_matching_session_tokens = 0;311 if (!session_tokens.empty()) {312 for (llama_token id : session_tokens) {313 if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {314 break;315 }316 n_matching_session_tokens++;317 }318 if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {319 LOG_INF("%s: using full prompt from session file\n", __func__);320 } else if (n_matching_session_tokens >= embd_inp.size()) {321 LOG_INF("%s: session file has exact match for prompt!\n", __func__);322 } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {323 LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",324 __func__, n_matching_session_tokens, embd_inp.size());325 } else {326 LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",327 __func__, n_matching_session_tokens, embd_inp.size());328 }329 330 // remove any "future" tokens that we might have inherited from the previous session331 llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);332 }333 334 LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",335 embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());336 337 // if we will use the cache for the full prompt without reaching the end of the cache, force338 // reevaluation of the last token to recalculate the cached logits339 if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {340 LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);341 342 session_tokens.resize(embd_inp.size() - 1);343 }344 345 // number of tokens to keep when resetting context346 if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {347 params.n_keep = (int)embd_inp.size();348 } else {349 params.n_keep += add_bos; // always keep the BOS token350 }351 352 if (params.conversation_mode) {353 params.interactive_first = true;354 }355 356 // enable interactive mode if interactive start is specified357 if (params.interactive_first) {358 params.interactive = true;359 }360 361 if (params.verbose_prompt) {362 LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());363 LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());364 for (int i = 0; i < (int) embd_inp.size(); i++) {365 LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());366 }367 368 if (params.n_keep > add_bos) {369 LOG_INF("%s: static prompt based on n_keep: '", __func__);370 for (int i = 0; i < params.n_keep; i++) {371 LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());372 }373 LOG_CNT("'\n");374 }375 LOG_INF("\n");376 }377 378 // ctrl+C handling379 {380#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))381 struct sigaction sigint_action;382 sigint_action.sa_handler = sigint_handler;383 sigemptyset (&sigint_action.sa_mask);384 sigint_action.sa_flags = 0;385 sigaction(SIGINT, &sigint_action, NULL);386#elif defined (_WIN32)387 auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {388 return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;389 };390 SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);391#endif392 }393 394 if (params.interactive) {395 LOG_INF("%s: interactive mode on.\n", __func__);396 397 if (!params.antiprompt.empty()) {398 for (const auto & antiprompt : params.antiprompt) {399 LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());400 if (params.verbose_prompt) {401 auto tmp = common_tokenize(ctx, antiprompt, false, true);402 for (int i = 0; i < (int) tmp.size(); i++) {403 LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());404 }405 }406 }407 }408 409 if (params.input_prefix_bos) {410 LOG_INF("Input prefix with BOS\n");411 }412 413 if (!params.input_prefix.empty()) {414 LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());415 if (params.verbose_prompt) {416 auto tmp = common_tokenize(ctx, params.input_prefix, true, true);417 for (int i = 0; i < (int) tmp.size(); i++) {418 LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());419 }420 }421 }422 423 if (!params.input_suffix.empty()) {424 LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());425 if (params.verbose_prompt) {426 auto tmp = common_tokenize(ctx, params.input_suffix, false, true);427 for (int i = 0; i < (int) tmp.size(); i++) {428 LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());429 }430 }431 }432 }433 434 smpl = common_sampler_init(model, sparams);435 if (!smpl) {436 LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);437 return 1;438 }439 440 LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));441 LOG_INF("sampler params: \n%s\n", sparams.print().c_str());442 LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());443 444 LOG_INF("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);445 446 // group-attention state447 // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)448 int ga_i = 0;449 450 const int ga_n = params.grp_attn_n;451 const int ga_w = params.grp_attn_w;452 453 if (ga_n != 1) {454 GGML_ASSERT(ga_n > 0 && "grp_attn_n must be positive"); // NOLINT455 GGML_ASSERT(ga_w % ga_n == 0 && "grp_attn_w must be a multiple of grp_attn_n"); // NOLINT456 //GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT457 //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT458 LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);459 }460 LOG_INF("\n");461 462 if (params.interactive) {463 const char * control_message;464 if (params.multiline_input) {465 control_message = " - To return control to the AI, end your input with '\\'.\n"466 " - To return control without starting a new line, end your input with '/'.\n";467 } else {468 control_message = " - Press Return to return control to the AI.\n"469 " - To return control without starting a new line, end your input with '/'.\n"470 " - If you want to submit another line, end your input with '\\'.\n";471 }472 LOG_INF("== Running in interactive mode. ==\n");473#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)474 LOG_INF( " - Press Ctrl+C to interject at any time.\n");475#endif476 LOG_INF( "%s", control_message);477 if (params.conversation_mode && params.enable_chat_template && params.prompt.empty()) {478 LOG_INF( " - Using default system message. To change it, set a different value via -p PROMPT or -f FILE argument.\n");479 }480 LOG_INF("\n");481 482 is_interacting = params.interactive_first;483 }484 485 bool is_antiprompt = false;486 bool input_echo = true;487 bool display = true;488 bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();489 490 int n_past = 0;491 int n_remain = params.n_predict;492 int n_consumed = 0;493 int n_session_consumed = 0;494 495 std::vector<int> input_tokens; g_input_tokens = &input_tokens;496 std::vector<int> output_tokens; g_output_tokens = &output_tokens;497 std::ostringstream output_ss; g_output_ss = &output_ss;498 std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode499 500 // the first thing we will do is to output the prompt, so set color accordingly501 console::set_display(console::prompt);502 display = params.display_prompt;503 504 std::vector<llama_token> embd;505 506 // single-token antiprompts507 std::vector<llama_token> antiprompt_token;508 509 for (const std::string & antiprompt : params.antiprompt) {510 auto ids = ::common_tokenize(ctx, antiprompt, false, true);511 if (ids.size() == 1) {512 antiprompt_token.push_back(ids[0]);513 }514 }515 516 if (llama_model_has_encoder(model)) {517 int enc_input_size = embd_inp.size();518 llama_token * enc_input_buf = embd_inp.data();519 520 if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {521 LOG_ERR("%s : failed to eval\n", __func__);522 return 1;523 }524 525 llama_token decoder_start_token_id = llama_model_decoder_start_token(model);526 if (decoder_start_token_id == LLAMA_TOKEN_NULL) {527 decoder_start_token_id = llama_vocab_bos(vocab);528 }529 530 embd_inp.clear();531 embd_inp.push_back(decoder_start_token_id);532 }533 534 while ((n_remain != 0 && !is_antiprompt) || params.interactive) {535 // predict536 if (!embd.empty()) {537 // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via538 // --prompt or --file which uses the same value.539 int max_embd_size = n_ctx - 4;540 541 // Ensure the input doesn't exceed the context size by truncating embd if necessary.542 if ((int) embd.size() > max_embd_size) {543 const int skipped_tokens = (int) embd.size() - max_embd_size;544 embd.resize(max_embd_size);545 546 console::set_display(console::error);547 LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");548 console::set_display(console::reset);549 }550 551 if (ga_n == 1) {552 // infinite text generation via context shifting553 // if we run out of context:554 // - take the n_keep first tokens from the original prompt (via n_past)555 // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches556 557 if (n_past + (int) embd.size() >= n_ctx) {558 if (!params.ctx_shift){559 LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);560 break;561 }562 563 if (params.n_predict == -2) {564 LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);565 break;566 }567 568 const int n_left = n_past - params.n_keep;569 const int n_discard = n_left/2;570 571 LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",572 n_past, n_left, n_ctx, params.n_keep, n_discard);573 574 llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);575 llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);576 577 n_past -= n_discard;578 579 LOG_DBG("after swap: n_past = %d\n", n_past);580 581 LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());582 583 LOG_DBG("clear session path\n");584 path_session.clear();585 }586 } else {587 // context extension via Self-Extend588 while (n_past >= ga_i + ga_w) {589 const int ib = (ga_n*ga_i)/ga_w;590 const int bd = (ga_w/ga_n)*(ga_n - 1);591 const int dd = (ga_w/ga_n) - ib*bd - ga_w;592 593 LOG_DBG("\n");594 LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);595 LOG_DBG("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);596 LOG_DBG("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);597 598 llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd);599 llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);600 llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd);601 602 n_past -= bd;603 604 ga_i += ga_w/ga_n;605 606 LOG_DBG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);607 }608 }609 610 // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)611 if (n_session_consumed < (int) session_tokens.size()) {612 size_t i = 0;613 for ( ; i < embd.size(); i++) {614 if (embd[i] != session_tokens[n_session_consumed]) {615 session_tokens.resize(n_session_consumed);616 break;617 }618 619 n_past++;620 n_session_consumed++;621 622 if (n_session_consumed >= (int) session_tokens.size()) {623 ++i;624 break;625 }626 }627 if (i > 0) {628 embd.erase(embd.begin(), embd.begin() + i);629 }630 }631 632 for (int i = 0; i < (int) embd.size(); i += params.n_batch) {633 int n_eval = (int) embd.size() - i;634 if (n_eval > params.n_batch) {635 n_eval = params.n_batch;636 }637 638 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());639 640 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {641 LOG_ERR("%s : failed to eval\n", __func__);642 return 1;643 }644 645 n_past += n_eval;646 647 LOG_DBG("n_past = %d\n", n_past);648 // Display total tokens alongside total time649 if (params.n_print > 0 && n_past % params.n_print == 0) {650 LOG_DBG("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);651 }652 }653 654 if (!embd.empty() && !path_session.empty()) {655 session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());656 n_session_consumed = session_tokens.size();657 }658 }659 660 embd.clear();661 662 if ((int) embd_inp.size() <= n_consumed && !is_interacting) {663 // optionally save the session on first sample (for faster prompt loading next time)664 if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {665 need_to_save_session = false;666 llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());667 668 LOG_DBG("saved session to %s\n", path_session.c_str());669 }670 671 const llama_token id = common_sampler_sample(smpl, ctx, -1);672 673 common_sampler_accept(smpl, id, /* accept_grammar= */ true);674 675 // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());676 677 embd.push_back(id);678 679 // echo this to console680 input_echo = true;681 682 // decrement remaining sampling budget683 --n_remain;684 685 LOG_DBG("n_remain: %d\n", n_remain);686 } else {687 // some user input remains from prompt or interaction, forward it to processing688 LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);689 while ((int) embd_inp.size() > n_consumed) {690 embd.push_back(embd_inp[n_consumed]);691 692 // push the prompt in the sampling context in order to apply repetition penalties later693 // for the prompt, we don't apply grammar rules694 common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);695 696 ++n_consumed;697 if ((int) embd.size() >= params.n_batch) {698 break;699 }700 }701 }702 703 // display text704 if (input_echo && display) {705 for (auto id : embd) {706 const std::string token_str = common_token_to_piece(ctx, id, params.special);707 708 // Console/Stream Output709 LOG("%s", token_str.c_str());710 711 // Record Displayed Tokens To Log712 // Note: Generated tokens are created one by one hence this check713 if (embd.size() > 1) {714 // Incoming Requested Tokens715 input_tokens.push_back(id);716 } else {717 // Outgoing Generated Tokens718 output_tokens.push_back(id);719 output_ss << token_str;720 }721 }722 }723 724 // reset color to default if there is no pending user input725 if (input_echo && (int) embd_inp.size() == n_consumed) {726 console::set_display(console::reset);727 display = true;728 }729 730 // if not currently processing queued inputs;731 if ((int) embd_inp.size() <= n_consumed) {732 // check for reverse prompt in the last n_prev tokens733 if (!params.antiprompt.empty()) {734 const int n_prev = 32;735 const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);736 737 is_antiprompt = false;738 // Check if each of the reverse prompts appears at the end of the output.739 // If we're not running interactively, the reverse prompt might be tokenized with some following characters740 // so we'll compensate for that by widening the search window a bit.741 for (std::string & antiprompt : params.antiprompt) {742 size_t extra_padding = params.interactive ? 0 : 2;743 size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)744 ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)745 : 0;746 747 if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {748 if (params.interactive) {749 is_interacting = true;750 }751 is_antiprompt = true;752 break;753 }754 }755 756 // check for reverse prompt using special tokens757 llama_token last_token = common_sampler_last(smpl);758 if (std::find(antiprompt_token.begin(), antiprompt_token.end(), last_token) != antiprompt_token.end()) {759 if (params.interactive) {760 is_interacting = true;761 }762 is_antiprompt = true;763 }764 765 if (is_antiprompt) {766 LOG_DBG("found antiprompt: %s\n", last_output.c_str());767 }768 }769 770 // deal with end of generation tokens in interactive mode771 if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {772 LOG_DBG("found an EOG token\n");773 774 if (params.interactive) {775 if (!params.antiprompt.empty()) {776 // tokenize and inject first reverse prompt777 const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);778 embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());779 is_antiprompt = true;780 }781 782 if (params.enable_chat_template) {783 chat_add_and_format("assistant", assistant_ss.str());784 }785 is_interacting = true;786 LOG("\n");787 }788 }789 790 // if current token is not EOG, we add it to current assistant message791 if (params.conversation_mode) {792 const auto id = common_sampler_last(smpl);793 assistant_ss << common_token_to_piece(ctx, id, false);794 }795 796 if (n_past > 0 && is_interacting) {797 LOG_DBG("waiting for user input\n");798 799 if (params.conversation_mode) {800 LOG("\n> ");801 }802 803 if (params.input_prefix_bos) {804 LOG_DBG("adding input prefix BOS token\n");805 embd_inp.push_back(llama_vocab_bos(vocab));806 }807 808 std::string buffer;809 if (!params.input_prefix.empty() && !params.conversation_mode) {810 LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());811 LOG("%s", params.input_prefix.c_str());812 }813 814 // color user input only815 console::set_display(console::user_input);816 display = params.display_prompt;817 818 std::string line;819 bool another_line = true;820 do {821 another_line = console::readline(line, params.multiline_input);822 buffer += line;823 } while (another_line);824 825 // done taking input, reset color826 console::set_display(console::reset);827 display = true;828 829 // Add tokens to embd only if the input buffer is non-empty830 // Entering a empty line lets the user pass control back831 if (buffer.length() > 1) {832 // append input suffix if any833 if (!params.input_suffix.empty() && !params.conversation_mode) {834 LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());835 LOG("%s", params.input_suffix.c_str());836 }837 838 LOG_DBG("buffer: '%s'\n", buffer.c_str());839 840 const size_t original_size = embd_inp.size();841 842 if (params.escape) {843 string_process_escapes(buffer);844 }845 846 bool format_chat = params.conversation_mode && params.enable_chat_template;847 std::string user_inp = format_chat848 ? chat_add_and_format("user", std::move(buffer))849 : std::move(buffer);850 // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)851 const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);852 const auto line_inp = common_tokenize(ctx, user_inp, false, format_chat);853 const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);854 855 LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());856 857 // if user stop generation mid-way, we must add EOT to finish model's last response858 if (need_insert_eot && format_chat) {859 llama_token eot = llama_vocab_eot(vocab);860 embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);861 need_insert_eot = false;862 }863 864 embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());865 embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());866 embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());867 868 for (size_t i = original_size; i < embd_inp.size(); ++i) {869 const llama_token token = embd_inp[i];870 output_tokens.push_back(token);871 output_ss << common_token_to_piece(ctx, token);872 }873 874 // reset assistant message875 assistant_ss.str("");876 877 n_remain -= line_inp.size();878 LOG_DBG("n_remain: %d\n", n_remain);879 } else {880 LOG_DBG("empty line, passing control back\n");881 }882 883 input_echo = false; // do not echo this again884 }885 886 if (n_past > 0) {887 if (is_interacting) {888 common_sampler_reset(smpl);889 }890 is_interacting = false;891 }892 }893 894 // end of generation895 if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {896 LOG(" [end of text]\n");897 break;898 }899 900 // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.901 // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).902 if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {903 n_remain = params.n_predict;904 is_interacting = true;905 }906 }907 908 if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {909 LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());910 llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());911 }912 913 LOG("\n\n");914 common_perf_print(ctx, smpl);915 916 common_sampler_free(smpl);917 918 llama_backend_free();919 920 ggml_threadpool_free_fn(threadpool);921 ggml_threadpool_free_fn(threadpool_batch);922 923 return 0;924}925