Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-20 01:31:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master' into xsn/server_tts_2

Browse Source
This commit is contained in:
Xuan Son Nguyen
2026-08-18 01:36:40 +02:00
parent cf5d754e11 01818e4956
commit 4daa33ddb0
1148 changed files with 51562 additions and 15901 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/tts/tts.cpp
+15 -7
View File
@@ -120,6 +120,7 @@ int main(int argc, char ** argv) {
inp.lang = params.tts_lang.c_str();
inp.top_k = params.sampling.top_k;
inp.top_p = params.sampling.top_p;
inp.seed = params.sampling.seed;
//
// stage 1: process prompt via backbone model, generate semantic representation
@@ -143,8 +144,7 @@ int main(int argc, char ** argv) {
}
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// note: some pipelines ignore this token and use the hidden state instead
auto sample_semantic_code = [&]() -> llama_token {
llama_token t = common_sampler_sample(smpl, lctx, -1);
common_sampler_accept(smpl, t, true);
@@ -159,19 +159,24 @@ int main(int argc, char ** argv) {
tts_timings timings;
const int64_t t_gen_start_us = ggml_time_us();
for (; n_frames < max_new && !llama_vocab_is_eog(vocab, sampled); n_frames++) {
bool stop = false;
while (!stop && n_frames < max_new) {
const float * h_next = nullptr;
// stage 2+3: semantic --> acoustic details --> audio waveform
// step_gen() runs both stages and returns new h_state for next step
if (gen.step_gen(sampled, h_state, &h_next) != 0) {
if (gen.step_gen(sampled, h_state, &h_next, &stop) != 0) {
LOG_ERR("step_gen failed at frame %d\n", n_frames);
return 1;
}
if (!h_next) {
break; // stopped without generating a frame
}
n_frames++;
h_state = h_next;
sampled = sample_semantic_code();
timings.report(n_frames + 1);
timings.report(n_frames);
}
const double t_gen_s = (ggml_time_us() - t_gen_start_us) / 1e6;
@@ -179,17 +184,20 @@ int main(int argc, char ** argv) {
const char * data = nullptr;
size_t data_len = 0;
int64_t n_samples = 0;
const int64_t t_wav_start_us = ggml_time_us();
if (gen.get_output(&sample_rate, &data, &data_len, &n_samples) != 0) {
LOG_ERR("get_output failed\n");
return 1;
}
const double t_wav_s = (ggml_time_us() - t_wav_start_us) / 1e6;
LOG_INF("generated %d frames, %zu bytes of WAV audio (%d Hz)\n", n_frames, data_len, sample_rate);
const double t_prompt_s = (t_gen_start_us - t_prompt_start_us) / 1e6;
const double t_total_s = t_prompt_s + t_gen_s;
const double t_total_s = t_prompt_s + t_gen_s + t_wav_s;
const double audio_s = sample_rate > 0 ? (double) n_samples / sample_rate : 0.0;
LOG_INF("timings: prompt eval %.2fs + generation %.2fs = total %.2fs\n", t_prompt_s, t_gen_s, t_total_s);
LOG_INF("timings: prompt eval %.2fs + generation %.2fs + vocoder %.2fs = total %.2fs\n",
t_prompt_s, t_gen_s, t_wav_s, t_total_s);
LOG_INF(" output audio = %.2fs (audio time = %.2fx process time)\n", audio_s, t_total_s > 0 ? audio_s / t_total_s : 0.0);
FILE * f = fopen(params.out_file.c_str(), "wb");
if (!f) {