llama.cpp/examples/main/main.cpp

#include "common.h"

#include "console.h"
#include "llama.h"
#include "build-info.h"
#include "grammar-parser.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 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 = get_sortable_timestamp();

    const bool success = 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));
    dump_non_result_info_yaml(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");

    dump_string_yaml_multiline(logfile, "output", output.c_str());
    dump_vector_int_yaml(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) {
            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

int main(int argc, char ** argv) {
    gpt_params params;
    g_params = &params;

    if (!gpt_params_parse(argc, argv, params)) {
        return 1;
    }

#ifndef LOG_DISABLE_LOGS
    log_set_target(log_filename_generator("main", "log"));
    LOG_TEE("Log start\n");
    log_dump_cmdline(argc, argv);
#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.perplexity) {
        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.rope_freq_base != 10000.0) {
        LOG_TEE("%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
    }

    if (params.rope_freq_scale != 1.0) {
        LOG_TEE("%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
    }

    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
    }

    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
        params.prompt = gpt_random_prompt(rng);
    }

    LOG("%s: llama backend init\n", __func__);
    llama_backend_init(params.numa);

    llama_model * model;
    llama_context * ctx;
    llama_context * ctx_guidance = NULL;
    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 (params.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(ctx);
    if (params.n_ctx > n_ctx_train) {
        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
                __func__, n_ctx_train, params.n_ctx);
    } else if (params.n_ctx < 8) {
        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
        params.n_ctx = 8;
    }

    // print system information
    {
        LOG_TEE("\n");
        LOG_TEE("system_info: n_threads = %d / %d | %s\n",
                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
    }

    // export the cgraph and exit
    if (params.export_cgraph) {
        llama_eval_export(ctx, "llama.ggml");
        llama_free(ctx);
        llama_free_model(model);

        return 0;
    }

    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());

        // fopen to check for existing session
        FILE * fp = std::fopen(path_session.c_str(), "rb");
        if (fp != NULL) {
            std::fclose(fp);

            session_tokens.resize(params.n_ctx);
            size_t n_token_count_out = 0;
            if (!llama_load_session_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);
            llama_set_rng_seed(ctx, params.seed);

            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
        } else {
            LOG_TEE("%s: session file does not exist, will create\n", __func__);
        }
    }

    const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
    LOG("add_bos: %d\n", add_bos);

    std::vector<llama_token> embd_inp;

    if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
        LOG("tokenize the prompt\n");
        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
    } else {
        LOG("use session tokens\n");
        embd_inp = session_tokens;
    }

    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));

    // Should not run without any tokens
    if (embd_inp.empty()) {
        embd_inp.push_back(llama_token_bos(ctx));
        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
    }

    // 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(params.cfg_negative_prompt));

        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));

        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));

        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));
    }

    const int n_ctx = llama_n_ctx(ctx);
    LOG("n_ctx: %d\n", n_ctx);

    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());
        }
    }

    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 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.instruct) {
        params.n_keep = (int)embd_inp.size();
    }

    // prefix & suffix for instruct mode
    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);

    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));

    // in instruct mode, we inject a prefix and a suffix to each input by the user
    if (params.instruct) {
        params.interactive_first = true;
        params.antiprompt.push_back("### Instruction:\n\n");
    }

    // 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__, params.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 > 0) {
        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");
    }

    if (params.interactive) {
#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

        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.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.input_suffix.empty()) {
            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
        }
    }
    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
    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);
    LOG_TEE("\n\n");

    struct llama_grammar * grammar = NULL;
    grammar_parser::parse_state parsed_grammar;

    if (!params.grammar.empty()) {
        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
        // will be empty (default) if there are parse errors
        if (parsed_grammar.rules.empty()) {
            return 1;
        }
        LOG_TEE("%s: grammar:\n", __func__);
        grammar_parser::print_grammar(stderr, parsed_grammar);
        LOG_TEE("\n");

        {
            auto it = params.logit_bias.find(llama_token_eos(ctx));
            if (it != params.logit_bias.end() && it->second == -INFINITY) {
                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
            }
        }

        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
        grammar = llama_grammar_init(
            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
    }

    // TODO: replace with ring-buffer
    std::vector<llama_token> last_tokens(n_ctx);
    std::fill(last_tokens.begin(), last_tokens.end(), 0);

    if (params.interactive) {
        const char *control_message;
        if (params.multiline_input) {
            control_message = " - To return control to LLaMa, 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 LLaMa.\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 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;

    // the first thing we will do is to output the prompt, so set color accordingly
    console::set_display(console::prompt);

    std::vector<llama_token> embd;
    std::vector<llama_token> embd_guidance;

    const int n_vocab = llama_n_vocab(ctx);

    std::vector<llama_token_data> candidates;
    candidates.reserve(n_vocab);

    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);
            }

            // infinite text generation via context swapping
            // 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;
                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep);

                // always keep the first token - BOS
                n_past          = std::max(1, params.n_keep);
                n_past_guidance = std::max(1, params.n_keep + guidance_offset);

                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);

                // insert n_left/2 tokens at the start of embd from last_tokens
                embd.insert(embd.begin(), last_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_tokens.end() - embd.size());

                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));

                LOG("clear session path\n");
                path_session.clear();
            }

            // 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));
                } 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_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
                        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));

                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
                    LOG_TEE("%s : failed to eval\n", __func__);
                    return 1;
                }

                n_past += n_eval;

                LOG("n_past = %d\n", n_past);
            }

            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_save_session_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_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);

            last_tokens.erase(last_tokens.begin());
            last_tokens.push_back(id);

            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));

            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]);
                last_tokens.erase(last_tokens.begin());
                last_tokens.push_back(embd_inp[n_consumed]);
                ++n_consumed;
                if ((int) embd.size() >= params.n_batch) {
                    break;
                }
            }
        }

        // display text
        if (input_echo) {
            for (auto id : embd) {
                const std::string token_str = llama_token_to_piece(ctx, id);
                printf("%s", token_str.c_str());

                if (embd.size() > 1) {
                    input_tokens.push_back(id);
                } else {
                    output_tokens.push_back(id);
                    output_ss << token_str;
                }
            }
            fflush(stdout);
        }
        // reset color to default if we there is no pending user input
        if (input_echo && (int) embd_inp.size() == n_consumed) {
            console::set_display(console::reset);
        }

        // if not currently processing queued inputs;
        if ((int) embd_inp.size() <= n_consumed) {
            // check for reverse prompt
            if (!params.antiprompt.empty()) {
                std::string last_output;
                for (auto id : last_tokens) {
                    last_output += llama_token_to_piece(ctx, id);
                }

                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;
                            console::set_display(console::user_input);
                        }
                        is_antiprompt = true;
                        fflush(stdout);
                        break;
                    }
                }

                if (is_antiprompt) {
                    LOG("found antiprompt: %s\n", last_output.c_str());
                }
            }

            // deal with end of text token in interactive mode
            if (last_tokens.back() == llama_token_eos(ctx)) {
                LOG("found EOS 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);
                        embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
                        is_antiprompt = true;
                    }

                    is_interacting = true;
                    printf("\n");
                    console::set_display(console::user_input);
                    fflush(stdout);
                } else if (params.instruct) {
                    is_interacting = true;
                }
            }

            if (n_past > 0 && is_interacting) {
                LOG("waiting for user input\n");

                if (params.instruct) {
                    printf("\n> ");
                }

                if (params.input_prefix_bos) {
                    LOG("adding input prefix BOS token\n");
                    embd_inp.push_back(llama_token_bos(ctx));
                }

                std::string buffer;
                if (!params.input_prefix.empty()) {
                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                    buffer += params.input_prefix;
                    printf("%s", buffer.c_str());
                }

                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);

                // 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()) {
                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                        buffer += params.input_suffix;
                        printf("%s", params.input_suffix.c_str());
                    }

                    LOG("buffer: '%s'\n", buffer.c_str());

                    const size_t original_size = embd_inp.size();

                    // instruct mode: insert instruction prefix
                    if (params.instruct && !is_antiprompt) {
                        LOG("inserting instruction prefix\n");
                        n_consumed = embd_inp.size();
                        embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                    }

                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));

                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());

                    // instruct mode: insert response suffix
                    if (params.instruct) {
                        LOG("inserting instruction suffix\n");
                        embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_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);
                    }

                    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) {
                    // reset grammar state if we're restarting generation
                    if (grammar != NULL) {
                        llama_grammar_free(grammar);

                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
                        grammar = llama_grammar_init(
                            grammar_rules.data(), grammar_rules.size(),
                            parsed_grammar.symbol_ids.at("root"));
                    }
                }
                is_interacting = false;
            }
        }

        // end of text token
        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || 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_save_session_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);

    if (grammar != NULL) {
        llama_grammar_free(grammar);
    }
    llama_backend_free();

#ifndef LOG_DISABLE_LOGS
    LOG_TEE("Log end\n")
#endif // LOG_DISABLE_LOGS

    return 0;
}