hey! this is a static git repo viewer. commit diffs, file history, older commits and other revisions won't load without javascript.

you are still able to view source files on branch/tag tips, but the experience is degraded.

char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Niklas Wenzelcommon: apply CPU parameters across tools (#27026)a94d563ed2026-08-13

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.h"
8
9#include <clocale>
10#include <cstdio>
11#include <cstring>
12#include <ctime>
13#include <fstream>
14#include <iostream>
15#include <sstream>
16#include <string>
17#include <vector>
18
19#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
20#include <signal.h>
21#include <unistd.h>
22#elif defined (_WIN32)
23#define WIN32_LEAN_AND_MEAN
24#ifndef NOMINMAX
25#define NOMINMAX
26#endif
27#include <windows.h>
28#include <signal.h>
29#endif
30
31#if defined(_MSC_VER)
32#pragma warning(disable: 4244 4267) // possible loss of data
33#endif
34
35static llama_context           ** g_ctx;
36static common_sampler          ** g_smpl;
37static common_params            * g_params;
38static bool is_interacting  = false;
39static bool need_insert_eot = false;
40
41static void print_usage(int argc, char ** argv) {
42    (void) argc;
43
44    LOG("\nexample usage:\n");
45    LOG("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128 -no-cnv\n", argv[0]);
46    LOG("\n  chat (conversation): %s -m your_model.gguf -sys \"You are a helpful assistant\"\n", argv[0]);
47    LOG("\n");
48}
49
50static bool file_exists(const std::string & path) {
51    std::ifstream f(path.c_str());
52    return f.good();
53}
54
55static bool file_is_empty(const std::string & path) {
56    std::ifstream f;
57    f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
58    f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
59    return f.tellg() == 0;
60}
61
62#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
63static void sigint_handler(int signo) {
64    if (signo == SIGINT) {
65        if (!is_interacting && g_params->interactive) {
66            is_interacting  = true;
67            need_insert_eot = true;
68        } else {
69            console::cleanup();
70            LOG("\n");
71            common_perf_print(*g_ctx, *g_smpl);
72
73            // make sure all logs are flushed
74            LOG("Interrupted by user\n");
75            common_log_pause(common_log_main());
76
77            _exit(130);
78        }
79    }
80}
81#endif
82
83// satisfies -Wmissing-declarations
84int llama_completion(int argc, char ** argv);
85
86int llama_completion(int argc, char ** argv) {
87    std::setlocale(LC_NUMERIC, "C");
88
89    common_params params;
90    g_params = &params;
91
92    common_init();
93
94    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMPLETION, print_usage)) {
95        return 1;
96    }
97
98    auto & sparams = params.sampling;
99
100    // save choice to use color for later
101    // (note for later: this is a slightly awkward choice)
102    console::init(params.simple_io, params.use_color);
103    atexit([]() { console::cleanup(); });
104
105    if (params.embedding) {
106        LOG_ERR("************\n");
107        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
108        LOG_ERR("************\n\n");
109
110        return 0;
111    }
112
113    if (params.n_ctx != 0 && params.n_ctx < 8) {
114        LOG_WRN("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
115        params.n_ctx = 8;
116    }
117
118    if (params.rope_freq_base != 0.0) {
119        LOG_WRN("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
120    }
121
122    if (params.rope_freq_scale != 0.0) {
123        LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
124    }
125
126    LOG_INF("%s: llama backend init\n", __func__);
127
128    llama_backend_init();
129    llama_numa_init(params.numa);
130
131    llama_model * model = nullptr;
132    llama_context * ctx = nullptr;
133    common_sampler * smpl = nullptr;
134
135    g_ctx = &ctx;
136    g_smpl = &smpl;
137
138    std::vector<common_chat_msg> chat_msgs;
139
140    // load the model and apply lora adapter, if any
141    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
142
143    auto llama_init = common_init_from_params(params);
144
145    ctx   = llama_init->context();
146    model = llama_init->model();
147    smpl  = llama_init->sampler(0);
148
149    if (ctx == NULL) {
150        LOG_ERR("%s: error: unable to create context\n", __func__);
151        return 1;
152    }
153
154    llama_memory_t mem = llama_get_memory(ctx);
155    const llama_vocab * vocab = llama_model_get_vocab(model);
156
157    // note: the time for chat template initialization is not negligible:
158    auto chat_templates = common_chat_templates_init(model, params.chat_template);
159
160    // start measuring performance timings from here
161    llama_perf_context_reset(ctx);
162
163    const int n_ctx_train = llama_model_n_ctx_train(model);
164    const int n_ctx = llama_n_ctx(ctx);
165
166    if (n_ctx > n_ctx_train) {
167        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
168    }
169
170    // auto enable conversation mode if chat template is available
171    const bool has_chat_template = common_chat_templates_was_explicit(chat_templates.get());
172    if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
173        if (has_chat_template) {
174            LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
175            params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
176        } else {
177            params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
178        }
179    }
180
181    // in case user force-activate conversation mode (via -cnv) without proper chat template, we show a warning
182    if (params.conversation_mode && !has_chat_template) {
183        LOG_WRN("%s: chat template is not available or is not supported. This may cause the model to output suboptimal responses\n", __func__);
184    }
185
186    // print chat template example in conversation mode
187    if (params.conversation_mode) {
188        if (params.enable_chat_template) {
189            if (!params.prompt.empty() && params.system_prompt.empty()) {
190                LOG_WRN("*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?\n");
191            }
192
193            LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs).c_str());
194        } else {
195            LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
196        }
197    }
198
199    // print system information
200    {
201        LOG_INF("\n");
202        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
203        LOG_INF("\n");
204    }
205
206    std::string path_session = params.path_prompt_cache;
207    std::vector<llama_token> session_tokens;
208
209    if (!path_session.empty()) {
210        LOG_INF("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
211        if (!file_exists(path_session)) {
212            LOG_INF("%s: session file does not exist, will create.\n", __func__);
213        } else if (file_is_empty(path_session)) {
214            LOG_INF("%s: The session file is empty. A new session will be initialized.\n", __func__);
215        } else {
216            // The file exists and is not empty
217            session_tokens.resize(n_ctx);
218            size_t n_token_count_out = 0;
219            if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
220                LOG_ERR("%s: failed to load session file '%s'\n", __func__, path_session.c_str());
221                return 1;
222            }
223            session_tokens.resize(n_token_count_out);
224            LOG_INF("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
225        }
226    }
227
228    const bool add_bos = llama_vocab_get_add_bos(vocab) && !params.use_jinja;
229    if (!llama_model_has_encoder(model)) {
230        GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
231    }
232
233    LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
234
235    std::vector<llama_token> embd_inp;
236
237    bool waiting_for_first_input = false;
238    auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
239        common_chat_msg new_msg;
240        new_msg.role = role;
241        new_msg.content = content;
242        auto formatted = common_chat_format_single(chat_templates.get(), chat_msgs, new_msg, role == "user", g_params->use_jinja);
243        chat_msgs.push_back(new_msg);
244        LOG_DBG("formatted: '%s'\n", formatted.c_str());
245        return formatted;
246    };
247
248    std::string prompt;
249    {
250        if (params.conversation_mode && params.enable_chat_template) {
251            if (!params.system_prompt.empty()) {
252                // format the system prompt (will use template default if empty)
253                chat_add_and_format("system", params.system_prompt);
254            }
255
256            if (!params.prompt.empty()) {
257                // format and append the user prompt
258                chat_add_and_format("user", params.prompt);
259            } else {
260                waiting_for_first_input = true;
261            }
262
263            if (!params.system_prompt.empty() || !params.prompt.empty()) {
264                common_chat_templates_inputs inputs;
265                inputs.use_jinja = g_params->use_jinja;
266                inputs.messages = chat_msgs;
267                inputs.add_generation_prompt = !params.prompt.empty();
268                inputs.force_pure_content = params.force_pure_content_parser;
269
270                prompt = common_chat_templates_apply(chat_templates.get(), inputs).prompt;
271            }
272        } else {
273            // otherwise use the prompt as is
274            prompt = params.prompt;
275        }
276
277        if (params.interactive_first || !prompt.empty() || session_tokens.empty()) {
278            LOG_DBG("tokenize the prompt\n");
279            embd_inp = common_tokenize(ctx, prompt, true, true);
280        } else {
281            LOG_DBG("use session tokens\n");
282            embd_inp = session_tokens;
283        }
284
285        LOG_DBG("prompt: \"%s\"\n", prompt.c_str());
286        LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
287    }
288
289    // Should not run without any tokens
290    if (!waiting_for_first_input && embd_inp.empty()) {
291        if (add_bos) {
292            embd_inp.push_back(llama_vocab_bos(vocab));
293            LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
294        } else {
295            LOG_ERR("input is empty\n");
296            return -1;
297        }
298    }
299
300    // Tokenize negative prompt
301    if ((int) embd_inp.size() > n_ctx - 4) {
302        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
303        return 1;
304    }
305
306    bool session_do_save = false;
307
308    {
309        size_t n_match = 0;
310
311        if (!session_tokens.empty()) {
312            for (llama_token id : session_tokens) {
313                if (n_match >= embd_inp.size() || id != embd_inp[n_match]) {
314                    break;
315                }
316                n_match++;
317            }
318            if (params.prompt.empty() && n_match == embd_inp.size()) {
319                LOG_INF("%s: using full prompt from session file\n", __func__);
320            } else if (n_match >= embd_inp.size()) {
321                LOG_INF("%s: session file has exact match for prompt!\n", __func__);
322            } else if (n_match < (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_match, embd_inp.size());
325            } else {
326                LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
327                        __func__, n_match, embd_inp.size());
328            }
329
330            // remove any "future" tokens that we might have inherited from the previous session
331            if (session_tokens.size() > n_match) {
332                llama_pos pos = n_match > 0 ? (llama_pos)(n_match - 1) : 0;
333                if (!llama_memory_seq_rm(mem, -1, pos, -1)) {
334                    LOG_WRN("%s: unable to reuse common prefix (for example, when the memory is recurrent)\n", __func__);
335                    llama_memory_clear(mem, true);
336                    session_tokens.clear();
337                    n_match = 0;
338                } else {
339                    session_tokens.resize(n_match);
340                }
341            }
342        }
343
344        session_do_save = !path_session.empty() && n_match < embd_inp.size() && !params.prompt_cache_ro;
345
346        // Logits are not stored as part of the session state so we need to
347        // "replay" the last token to get logits for sampling.
348        if (!session_tokens.empty() && n_match > 0 && n_match == session_tokens.size()) {
349            if (!common_replay_last_token(ctx, session_tokens.back(), n_match - 1)) {
350                return 1;
351            }
352
353            session_do_save = false;
354            LOG_INF("%s: replayed last token from session\n", __func__);
355        }
356    }
357
358    // number of tokens to keep when resetting context
359    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
360        params.n_keep = (int)embd_inp.size();
361    } else {
362        params.n_keep += add_bos; // always keep the BOS token
363    }
364
365    if (params.conversation_mode) {
366        if (params.single_turn && !params.prompt.empty()) {
367            params.interactive = false;
368            params.interactive_first = false;
369        } else {
370            params.interactive_first = true;
371        }
372    }
373
374    // enable interactive mode if interactive start is specified
375    if (params.interactive_first) {
376        params.interactive = true;
377    }
378
379    if (params.verbose_prompt) {
380        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
381        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
382        for (int i = 0; i < (int) embd_inp.size(); i++) {
383            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
384        }
385
386        if (params.n_keep > add_bos) {
387            LOG_INF("%s: static prompt based on n_keep: '", __func__);
388            for (int i = 0; i < params.n_keep; i++) {
389                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
390            }
391            LOG_CNT("'\n");
392        }
393        LOG_INF("\n");
394    }
395
396    // ctrl+C handling
397    {
398#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
399        struct sigaction sigint_action;
400        sigint_action.sa_handler = sigint_handler;
401        sigemptyset (&sigint_action.sa_mask);
402        sigint_action.sa_flags = 0;
403        sigaction(SIGINT, &sigint_action, NULL);
404#elif defined (_WIN32)
405        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
406            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
407        };
408        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
409#endif
410    }
411
412    if (params.interactive) {
413        LOG_INF("%s: interactive mode on.\n", __func__);
414
415        if (!params.antiprompt.empty()) {
416            for (const auto & antiprompt : params.antiprompt) {
417                LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());
418                if (params.verbose_prompt) {
419                    auto tmp = common_tokenize(ctx, antiprompt, false, true);
420                    for (int i = 0; i < (int) tmp.size(); i++) {
421                        LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
422                    }
423                }
424            }
425        }
426
427        if (params.input_prefix_bos) {
428            LOG_INF("Input prefix with BOS\n");
429        }
430
431        if (!params.input_prefix.empty()) {
432            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
433            if (params.verbose_prompt) {
434                auto tmp = common_tokenize(ctx, params.input_prefix, true, true);
435                for (int i = 0; i < (int) tmp.size(); i++) {
436                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
437                }
438            }
439        }
440
441        if (!params.input_suffix.empty()) {
442            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
443            if (params.verbose_prompt) {
444                auto tmp = common_tokenize(ctx, params.input_suffix, false, true);
445                for (int i = 0; i < (int) tmp.size(); i++) {
446                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
447                }
448            }
449        }
450    }
451
452    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
453    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
454    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
455
456    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);
457
458    // group-attention state
459    // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
460    int ga_i = 0;
461
462    const int ga_n = params.grp_attn_n;
463    const int ga_w = params.grp_attn_w;
464
465    if (ga_n != 1) {
466        GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
467        GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
468      //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
469      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
470        LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
471    }
472    LOG_INF("\n");
473
474    if (params.interactive) {
475        const char * control_message;
476        if (params.multiline_input) {
477            control_message = " - To return control to the AI, end your input with '\\'.\n"
478                              " - To return control without starting a new line, end your input with '/'.\n";
479        } else {
480            control_message = " - Press Return to return control to the AI.\n"
481                              " - To return control without starting a new line, end your input with '/'.\n"
482                              " - If you want to submit another line, end your input with '\\'.\n";
483        }
484        LOG_INF("== Running in interactive mode. ==\n");
485#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
486        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
487#endif
488        LOG_INF(       "%s", control_message);
489        if (params.conversation_mode && params.enable_chat_template && params.system_prompt.empty()) {
490            LOG_INF(   " - Not using system message. To change it, set a different value via -sys PROMPT\n");
491        }
492        LOG_INF("\n");
493
494        is_interacting = params.interactive_first;
495    }
496
497    bool is_antiprompt = false;
498    bool input_echo    = true;
499    bool display       = true;
500
501    int n_past             = 0;
502    int n_remain           = params.n_predict;
503    int n_consumed         = 0;
504    int n_session_consumed = 0;
505
506    std::vector<int>   input_tokens;
507    std::vector<int>   output_tokens;
508    std::ostringstream output_ss;
509    std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
510
511    // the first thing we will do is to output the prompt, so set color accordingly
512    console::set_display(DISPLAY_TYPE_PROMPT);
513    display = params.display_prompt;
514
515    std::vector<llama_token> embd;
516
517    // single-token antiprompts
518    std::vector<llama_token> antiprompt_token;
519
520    for (const std::string & antiprompt : params.antiprompt) {
521        auto ids = ::common_tokenize(ctx, antiprompt, false, true);
522        if (ids.size() == 1) {
523            antiprompt_token.push_back(ids[0]);
524        }
525    }
526
527    if (llama_model_has_encoder(model)) {
528        int enc_input_size = embd_inp.size();
529        llama_token * enc_input_buf = embd_inp.data();
530
531        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
532            LOG_ERR("%s : failed to eval\n", __func__);
533            return 1;
534        }
535
536        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
537        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
538            decoder_start_token_id = llama_vocab_bos(vocab);
539        }
540
541        embd_inp.clear();
542        embd_inp.push_back(decoder_start_token_id);
543    }
544
545    while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
546        // predict
547        if (!embd.empty()) {
548            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
549            // --prompt or --file which uses the same value.
550            int max_embd_size = n_ctx - 4;
551
552            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
553            if ((int) embd.size() > max_embd_size) {
554                const int skipped_tokens = (int) embd.size() - max_embd_size;
555                embd.resize(max_embd_size);
556
557                console::set_display(DISPLAY_TYPE_ERROR);
558                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
559                console::set_display(DISPLAY_TYPE_RESET);
560            }
561
562            if (ga_n == 1) {
563                // infinite text generation via context shifting
564                // if we run out of context:
565                // - take the n_keep first tokens from the original prompt (via n_past)
566                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
567
568                if (n_past + (int) embd.size() >= n_ctx) {
569                    if (!params.ctx_shift){
570                        LOG_WRN("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
571                        break;
572                    }
573
574                    if (params.n_predict == -2) {
575                        LOG_WRN("\n\n%s: context full and n_predict == %d => stopping\n", __func__, params.n_predict);
576                        break;
577                    }
578
579                    const int n_left    = n_past - params.n_keep;
580                    const int n_discard = n_left/2;
581
582                    LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
583                            n_past, n_left, n_ctx, params.n_keep, n_discard);
584
585                    llama_memory_seq_rm (mem, 0, params.n_keep            , params.n_keep + n_discard);
586                    llama_memory_seq_add(mem, 0, params.n_keep + n_discard, n_past, -n_discard);
587
588                    n_past -= n_discard;
589
590                    LOG_DBG("after swap: n_past = %d\n", n_past);
591
592                    LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
593
594                    LOG_DBG("clear session path\n");
595                    path_session.clear();
596                }
597            } else {
598                // context extension via Self-Extend
599                while (n_past >= ga_i + ga_w) {
600                    const int ib = (ga_n*ga_i)/ga_w;
601                    const int bd = (ga_w/ga_n)*(ga_n - 1);
602                    const int dd = (ga_w/ga_n) - ib*bd - ga_w;
603
604                    LOG_DBG("\n");
605                    LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
606                    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);
607                    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);
608
609                    llama_memory_seq_add(mem, 0, ga_i,                n_past,              ib*bd);
610                    llama_memory_seq_div(mem, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
611                    llama_memory_seq_add(mem, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
612
613                    n_past -= bd;
614
615                    ga_i += ga_w/ga_n;
616
617                    LOG_DBG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
618                }
619            }
620
621            // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
622            if (n_session_consumed < (int) session_tokens.size()) {
623                size_t i = 0;
624                for ( ; i < embd.size(); i++) {
625                    if (embd[i] != session_tokens[n_session_consumed]) {
626                        session_tokens.resize(n_session_consumed);
627                        break;
628                    }
629
630                    n_past++;
631                    n_session_consumed++;
632
633                    if (n_session_consumed >= (int) session_tokens.size()) {
634                        ++i;
635                        break;
636                    }
637                }
638                if (i > 0) {
639                    embd.erase(embd.begin(), embd.begin() + i);
640                }
641            }
642
643            if (!embd.empty()) {
644                const bool is_last_batch = (n_consumed >= (int) embd_inp.size());
645                const bool save_now = session_do_save && is_last_batch;
646                session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
647                if (!common_prompt_batch_decode(ctx, session_tokens, embd.size(), n_past, params.n_batch, path_session, save_now)) {
648                    return 1;
649                }
650                n_session_consumed += embd.size();
651                if (save_now) {
652                    session_do_save = false;
653                }
654
655                LOG_DBG("n_past = %d\n", n_past);
656
657                // Display total tokens alongside total time
658                if (params.n_print > 0 && n_past % params.n_print == 0) {
659                    LOG_DBG("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
660                }
661            }
662        }
663
664        embd.clear();
665
666        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
667
668            const llama_token id = common_sampler_sample(smpl, ctx, -1);
669
670            common_sampler_accept(smpl, id, /* accept_grammar= */ true);
671
672            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
673
674            embd.push_back(id);
675
676            if (params.conversation_mode && !waiting_for_first_input && !llama_vocab_is_eog(vocab, id)) {
677                assistant_ss << common_token_to_piece(ctx, id, false);
678            }
679
680            // echo this to console
681            input_echo = true;
682
683            // decrement remaining sampling budget
684            --n_remain;
685
686            LOG_DBG("n_remain: %d\n", n_remain);
687        } else {
688            // some user input remains from prompt or interaction, forward it to processing
689            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
690            while ((int) embd_inp.size() > n_consumed) {
691                embd.push_back(embd_inp[n_consumed]);
692
693                // push the prompt in the sampling context in order to apply repetition penalties later
694                // for the prompt, we don't apply grammar rules
695                common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
696
697                ++n_consumed;
698                if ((int) embd.size() == params.n_batch) {
699                    break;
700                }
701            }
702        }
703
704        // display text
705        if (input_echo && display) {
706            for (auto id : embd) {
707                const std::string token_str = common_token_to_piece(ctx, id, params.special);
708
709                // Console/Stream Output
710                LOG("%s", token_str.c_str());
711
712                // Record Displayed Tokens To Log
713                // Note: Generated tokens are created one by one hence this check
714                if (embd.size() > 1) {
715                    // Incoming Requested Tokens
716                    input_tokens.push_back(id);
717                } else {
718                    // Outgoing Generated Tokens
719                    output_tokens.push_back(id);
720                    output_ss << token_str;
721                }
722            }
723        }
724
725        // reset color to default if there is no pending user input
726        if (input_echo && (int) embd_inp.size() == n_consumed) {
727            console::set_display(DISPLAY_TYPE_RESET);
728            display = true;
729        }
730
731        // if not currently processing queued inputs;
732        if ((int) embd_inp.size() <= n_consumed) {
733            // check for reverse prompt in the last n_prev tokens
734            if (!params.antiprompt.empty()) {
735                const int n_prev = 32;
736                const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);
737
738                is_antiprompt = false;
739                // Check if each of the reverse prompts appears at the end of the output.
740                // If we're not running interactively, the reverse prompt might be tokenized with some following characters
741                // so we'll compensate for that by widening the search window a bit.
742                for (std::string & antiprompt : params.antiprompt) {
743                    size_t extra_padding = params.interactive ? 0 : 2;
744                    size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)
745                        ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
746                        : 0;
747
748                    if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
749                        if (params.interactive) {
750                            is_interacting = true;
751                        }
752                        is_antiprompt = true;
753                        break;
754                    }
755                }
756
757                // check for reverse prompt using special tokens
758                // avoid calling common_sampler_last() if last_output is empty
759                if (!last_output.empty()) {
760                    llama_token last_token = common_sampler_last(smpl);
761                    for (auto token : antiprompt_token) {
762                        if (token == last_token) {
763                            if (params.interactive) {
764                                is_interacting = true;
765                            }
766                            is_antiprompt = true;
767                            break;
768                        }
769                    }
770                }
771
772                if (is_antiprompt) {
773                    LOG_DBG("found antiprompt: %s\n", last_output.c_str());
774                }
775            }
776
777            // deal with end of generation tokens in interactive mode
778            if (!waiting_for_first_input && llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
779                LOG_DBG("found an EOG token\n");
780
781                if (params.interactive) {
782                    if (!params.antiprompt.empty()) {
783                        // tokenize and inject first reverse prompt
784                        const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);
785                        embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
786                        is_antiprompt = true;
787                    }
788
789                    if (params.enable_chat_template) {
790                        chat_add_and_format("assistant", assistant_ss.str());
791                    }
792                    is_interacting = true;
793                    LOG("\n");
794                }
795            }
796
797            if (params.conversation_mode && !waiting_for_first_input) {
798                if (!prompt.empty()) {
799                    prompt.clear();
800                    is_interacting = false;
801                }
802            }
803
804            if ((n_past > 0 || waiting_for_first_input) && is_interacting) {
805                LOG_DBG("waiting for user input\n");
806
807                if (params.conversation_mode) {
808                    LOG("\n> ");
809                }
810
811                if (params.input_prefix_bos) {
812                    LOG_DBG("adding input prefix BOS token\n");
813                    embd_inp.push_back(llama_vocab_bos(vocab));
814                }
815
816                std::string buffer;
817                if (!params.input_prefix.empty() && !params.conversation_mode) {
818                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
819                    LOG("%s", params.input_prefix.c_str());
820                }
821
822                // color user input only
823                console::set_display(DISPLAY_TYPE_USER_INPUT);
824                display = params.display_prompt;
825
826                std::string line;
827                bool another_line = true;
828                do {
829                    another_line = console::readline(line, params.multiline_input);
830                    buffer += line;
831                } while (another_line);
832
833                // done taking input, reset color
834                console::set_display(DISPLAY_TYPE_RESET);
835                display = true;
836
837                if (buffer.empty()) { // Ctrl+D on empty line exits
838                    LOG("EOF by user\n");
839                    break;
840                }
841
842                if (buffer.back() == '\n') {
843                    // Implement #587:
844                    // If the user wants the text to end in a newline,
845                    // this should be accomplished by explicitly adding a newline by using \ followed by return,
846                    // then returning control by pressing return again.
847                    buffer.pop_back();
848                }
849
850                if (buffer.empty()) { // Enter key on empty line lets the user pass control back
851                    LOG_DBG("empty line, passing control back\n");
852                } else { // Add tokens to embd only if the input buffer is non-empty
853                    // append input suffix if any
854                    if (!params.input_suffix.empty() && !params.conversation_mode) {
855                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
856                        LOG("%s", params.input_suffix.c_str());
857                    }
858
859                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
860
861                    const size_t original_size = embd_inp.size();
862
863                    if (params.escape) {
864                        string_process_escapes(buffer);
865                    }
866
867                    bool format_chat = params.conversation_mode && params.enable_chat_template;
868                    std::string user_inp = format_chat
869                        ? chat_add_and_format("user", std::move(buffer))
870                        : std::move(buffer);
871                    // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
872                    const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);
873                    const auto line_inp = common_tokenize(ctx, user_inp,            false, format_chat);
874                    const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);
875
876                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
877
878                    // if user stop generation mid-way, we must add EOT to finish model's last response
879                    if (need_insert_eot && format_chat) {
880                        llama_token eot = llama_vocab_eot(vocab);
881                        embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);
882                        need_insert_eot = false;
883                    }
884
885                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
886                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
887                    embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
888
889                    if (params.verbose_prompt) {
890                        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size() - original_size);
891                    }
892
893                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
894                        const llama_token token = embd_inp[i];
895                        const std::string token_str = common_token_to_piece(ctx, token);
896                        output_tokens.push_back(token);
897                        output_ss << token_str;
898
899                        if (params.verbose_prompt) {
900                            LOG_INF("%6d -> '%s'\n", token, token_str.c_str());
901                        }
902                    }
903
904                    // reset assistant message
905                    assistant_ss.str("");
906
907                    n_remain -= line_inp.size();
908                    LOG_DBG("n_remain: %d\n", n_remain);
909                }
910
911                input_echo = false; // do not echo this again
912            }
913
914            if (n_past > 0 || waiting_for_first_input) {
915                if (is_interacting) {
916                    common_sampler_reset(smpl);
917                }
918                is_interacting = false;
919
920                if (waiting_for_first_input && params.single_turn) {
921                    params.interactive = false;
922                    params.interactive_first = false;
923                }
924                waiting_for_first_input = false;
925            }
926        }
927
928        // end of generation
929        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {
930            LOG(" [end of text]\n");
931            break;
932        }
933
934        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
935        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
936        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
937            n_remain = params.n_predict;
938            is_interacting = true;
939        }
940    }
941
942    if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
943        LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
944        session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
945        llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
946        LOG_INF("saved final session to %s, n_tokens = %zu\n", path_session.data(), session_tokens.size());
947
948    }
949
950    LOG("\n\n");
951    common_perf_print(ctx, smpl);
952
953    llama_backend_free();
954
955    return 0;
956}