Felipe97/llama-cpp-compiled

1#include "arg.h"2#include "common.h"3#include "sampling.h"4#include "log.h"5#include "llama.h"6#include "mtmd.h"7#include "mtmd-helper.h"8 9#include <cstdio>10#include <cstring>11#include <string>12 13/**14 * Please note that this is NOT a production-ready binary.15 * It is a playground for trying TTS support in llama.cpp.16 * For contributors: please keep this code simple and easy to understand. Do not add unnecessary complexity. The goal is to have a simple CLI for testing TTS support.17 */18 19struct tts_timings {20    int64_t t_start_us = ggml_time_us();21    int64_t t_last_us  = t_start_us;22 23    void report(int n_frames) {24        const int64_t t_now_us = ggml_time_us();25        if (t_now_us - t_last_us < 2000000) {26            return;27        }28        t_last_us = t_now_us;29        const double t_elapsed_s = (t_now_us - t_start_us) / 1e6;30        const double fps = t_elapsed_s > 0 ? n_frames / t_elapsed_s : 0.0;31        LOG_INF("frames generated: %d, speed: %.2f frames/s\n", n_frames, fps);32    }33};34 35static void print_usage(int, char ** argv) {36    LOG("\nexample usage:\n");37    LOG("\n    %s -m backbone.gguf -mm mmproj.gguf -p \"text to speak\" -o output.wav", argv[0]);38    LOG("\n    %s -hf user/model -p \"text to speak\" -o output.wav\n", argv[0]);39    LOG("\nnote: --tts-lang and --tts-speaker-file may not be supported in all models");40    LOG("\n      use -n to limit the output length");41    LOG("\n      see tts/README.md for per-model usage notes");42    LOG("\n\n");43}44 45int main(int argc, char ** argv) {46    common_params params;47 48    common_init();49 50    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_TTS, print_usage)) {51        return 1;52    }53 54    mtmd_helper_log_set(common_log_default_callback, nullptr);55 56    if (params.prompt.empty()) {57        LOG_ERR("no prompt provided, use -p \"text\"\n");58        return 1;59    }60    if (params.mmproj.path.empty()) {61        LOG_ERR("no mmproj provided, use --mmproj\n");62        return 1;63    }64 65    // important: keep this file as generic as possible66    //            model-specific logic should be in mtmd-helper-gen or mtmd API67 68    // always enable embd, so that we can pass hidden states to the audio generation helper69    params.embedding = true;70 71    llama_backend_init();72    llama_numa_init(params.numa);73 74    //75    // load backbone model and mmproj76    //77 78    auto llama_init = common_init_from_params(params);79    llama_model    * model = llama_init->model();80    llama_context  * lctx  = llama_init->context();81    common_sampler * smpl  = llama_init->sampler(0);82    if (!model || !lctx) {83        LOG_ERR("failed to init model/context\n");84        return 1;85    }86 87    mtmd_context_params mtmd_params = mtmd_context_params_default();88    mtmd_params.use_gpu = params.mmproj_use_gpu;89    mtmd_params.device  = params.mmproj_device;90    mtmd::context_ptr mctx(mtmd_init_from_file(params.mmproj.path.c_str(), model, mtmd_params));91    if (!mctx) {92        LOG_ERR("failed to load mmproj %s\n", params.mmproj.path.c_str());93        return 1;94    }95    if (mtmd_gen_audio_get_info(mctx.get()).type == MTMD_GEN_AUDIO_TYPE_NONE) {96        LOG_ERR("mmproj does not support audio generation\n");97        return 1;98    }99 100    //101    // stage 0: process speaker reference file, if any102    //103 104    mtmd::bitmap_ptr speaker_bitmap;105    if (!params.tts_speaker_file.empty()) {106        auto wrapper = mtmd_helper_bitmap_init_from_file(mctx.get(), params.tts_speaker_file.c_str(), false, mtmd_helper_init_opt_default());107        if (!wrapper.bitmap) {108            LOG_ERR("failed to load speaker file %s\n", params.tts_speaker_file.c_str());109            return 1;110        }111        speaker_bitmap.reset(wrapper.bitmap);112    }113 114    mtmd_helper::gen_audio gen(lctx, mctx.get());115    mtmd_helper_gen_audio_inp inp{};116    inp.seq_id      = 0;117    inp.prompt      = params.prompt.c_str();118    inp.prompt_len  = params.prompt.size();119    inp.speaker_ref = speaker_bitmap.get();120    inp.lang        = params.tts_lang.c_str();121    inp.top_k       = params.sampling.top_k;122    inp.top_p       = params.sampling.top_p;123    inp.seed        = params.sampling.seed;124    inp.out_type    = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;125 126    //127    // stage 1: process prompt via backbone model, generate semantic representation128    //129 130    if (gen.set_input(&inp) != 0) {131        LOG_ERR("set_input failed\n");132        return 1;133    }134 135    const int64_t t_prompt_start_us = ggml_time_us();136 137    for (;;) {138        int32_t ret = gen.step_prompt(params.n_batch);139        if (ret < 0) {140            LOG_ERR("prompt processing failed\n");141            return 1;142        }143        if (ret == 0) {144            break;145        }146    }147 148    // note: some pipelines ignore this token and use the hidden state instead149    auto sample_semantic_code = [&]() -> llama_token {150        llama_token t = common_sampler_sample(smpl, lctx, -1);151        common_sampler_accept(smpl, t, true);152        return t;153    };154 155    const int max_new = params.n_predict > 0 ? params.n_predict : 512;156    int n_frames = 0;157    llama_token sampled = sample_semantic_code();158    const float * h_state = llama_get_embeddings_ith(lctx, -1);159 160    tts_timings timings;161    const int64_t t_gen_start_us = ggml_time_us();162 163    bool stop = false;164    while (!stop && n_frames < max_new) {165        const float * h_next = nullptr;166 167        // stage 2+3: semantic --> acoustic details --> audio waveform168        //            step_gen() runs both stages and returns new h_state for next step169        if (gen.step_gen(sampled, h_state, &h_next, &stop) != 0) {170            LOG_ERR("step_gen failed at frame %d\n", n_frames);171            return 1;172        }173        if (!h_next) {174            break; // stopped without generating a frame175        }176 177        n_frames++;178        h_state = h_next;179        sampled = sample_semantic_code();180        timings.report(n_frames);181    }182    const double t_gen_s = (ggml_time_us() - t_gen_start_us) / 1e6;183 184    int32_t      sample_rate = 0;185    const char * data        = nullptr;186    size_t       data_len    = 0;187    int64_t      n_samples   = 0;188    const int64_t t_wav_start_us = ggml_time_us();189    if (gen.get_output(&sample_rate, &data, &data_len, &n_samples) != 0) {190        LOG_ERR("get_output failed\n");191        return 1;192    }193    const double t_wav_s = (ggml_time_us() - t_wav_start_us) / 1e6;194 195    LOG_INF("generated %d frames, %zu bytes of WAV audio (%d Hz)\n", n_frames, data_len, sample_rate);196 197    const double t_prompt_s = (t_gen_start_us - t_prompt_start_us) / 1e6;198    const double t_total_s  = t_prompt_s + t_gen_s + t_wav_s;199    const double audio_s    = sample_rate > 0 ? (double) n_samples / sample_rate : 0.0;200    LOG_INF("timings: prompt eval %.2fs + generation %.2fs + vocoder %.2fs = total %.2fs\n",201            t_prompt_s, t_gen_s, t_wav_s, t_total_s);202    LOG_INF("         output audio = %.2fs (audio time = %.2fx process time)\n", audio_s, t_total_s > 0 ? audio_s / t_total_s : 0.0);203    FILE * f = fopen(params.out_file.c_str(), "wb");204    if (!f) {205        LOG_ERR("failed to open %s\n", params.out_file.c_str());206        return 1;207    }208    fwrite(data, 1, data_len, f);209    fclose(f);210    LOG_INF("wrote %s\n", params.out_file.c_str());211 212    llama_backend_free();213    return 0;214}215