Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-11 04:56:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

demo: use proper sampling

Browse Source
This commit is contained in:
Xuan Son Nguyen
2026-08-01 14:42:10 +02:00
parent 06aae58899
commit eae9c2c3f5
1 changed files with 23 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/tts/qwen3-tts.cpp
+23 -19
View File
@@ -7,6 +7,7 @@
#include "mtmd.h"
#include "mtmd-helper.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include <cstdio>
@@ -90,29 +91,31 @@ int main(int argc, char ** argv) {
if (gen.set_input(&inp) != 0) { LOG_ERR("set_input failed\n"); return 1; }
mtmd_bitmap_free(speaker_wrapper.bitmap);
// vocab landmarks: only the backbone-level sampling policy (which tokens
// are valid codec_0 candidates, and which one means EOS) stays here, since
// that's ordinary LLM sampling, not part of the audio-generation pipeline
llama_token codec_0_tok = LLAMA_TOKEN_NULL;
// codec_0 (backbone) EOS token: only this and the sampling policy below are
// ordinary LLM sampling concerns, kept out of the audio-generation helper
llama_token codec_eos_tok = LLAMA_TOKEN_NULL;
for (llama_token t = 0; t < llama_vocab_n_tokens(vocab); t++) {
const char * piece = llama_vocab_get_text(vocab, t);
if (!strcmp(piece, "<|codec_0|>")) codec_0_tok = t;
else if (!strcmp(piece, "<|codec_eos_token|>")) codec_eos_tok = t;
}
if (codec_0_tok == LLAMA_TOKEN_NULL || codec_eos_tok == LLAMA_TOKEN_NULL) {
LOG_ERR("missing codec special tokens in vocab\n");
return 1;
if (!strcmp(llama_vocab_get_text(vocab, t), "<|codec_eos_token|>")) { codec_eos_tok = t; break; }
}
if (codec_eos_tok == LLAMA_TOKEN_NULL) { LOG_ERR("missing codec eos token in vocab\n"); return 1; }
// reference defaults (qwen_tts Qwen3TTSForConditionalGeneration.generate()):
// top_k=50, top_p=1.0, temperature=0.9, repetition_penalty=1.05 over the whole
// generated history — without the penalty, runs degenerate into immediate
// no-speech (EOS) or endless non-terminating tails
common_params_sampling sparams;
sparams.top_k = 50;
sparams.top_p = 0.95f;
sparams.temp = 0.9f;
sparams.penalty_repeat = 1.05f;
sparams.penalty_last_n = -1;
common_sampler * smpl = common_sampler_init(model, sparams);
if (!smpl) { LOG_ERR("failed to init sampler\n"); return 1; }
auto sample_codec0 = [&]() -> llama_token {
const float * logits = llama_get_logits_ith(lctx, -1);
llama_token best = codec_eos_tok;
float bestv = logits[codec_eos_tok];
for (llama_token t = codec_0_tok; t < codec_0_tok + 2048; t++) {
if (logits[t] > bestv) { bestv = logits[t]; best = t; }
}
return best;
llama_token t = common_sampler_sample(smpl, lctx, -1);
common_sampler_accept(smpl, t, true);
return t;
};
int n_frames = 0;
@@ -120,10 +123,11 @@ int main(int argc, char ** argv) {
const float * h_state = llama_get_embeddings_ith(lctx, -1);
for (; n_frames < max_new && sampled != codec_eos_tok; n_frames++) {
const float * h_next = nullptr;
if (gen.step(sampled, h_state, &h_next) != 0) { LOG_ERR("step failed at frame %d\n", n_frames); return 1; }
if (gen.step(sampled, h_state, &h_next) != 0) { LOG_ERR("step failed at frame %d\n", n_frames); common_sampler_free(smpl); return 1; }
h_state = h_next;
sampled = sample_codec0();
}
common_sampler_free(smpl);
int32_t sample_rate = 0;
const char * data = nullptr;