Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-13 02:10:45 +02:00
Code Issues Packages Projects Releases Wiki Activity

rm gen_flow_temp

Browse Source
This commit is contained in:
Xuan Son Nguyen
2026-08-11 01:01:16 +02:00
parent 606375f8c3
commit f236e272f0
3 changed files with 6 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip-model.h
-3
View File
@@ -144,9 +144,6 @@ struct clip_hparams {
// threshold for the "out_eos_score" graph output
float gen_eos_threshold = 0.0f;
// default noise scale of a flow-matching decoder, see mtmd_gen_inp_default()
float gen_flow_temp = 0.0f;
// name of the weight variant, some pipelines tune themselves on it
std::string gen_model_variant;
tools/mtmd/clip.cpp
+2 -2
View File
@@ -1784,7 +1784,6 @@ struct clip_model_loader {
// flow_lm defaults, see pocket_tts/default_parameters.py
hparams.flow_n_step = 1;
hparams.gen_eos_threshold = -4.0f;
hparams.gen_flow_temp = 0.7f; // Config.default_temperature
} break;
case PROJECTOR_TYPE_PADDLEOCR:
{
@@ -4964,7 +4963,8 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
} else {
// flow matching starts from gaussian noise, std = sqrt(temp)
ggml_tensor * t = get_inp_tensor("inp_noise");
const float temp = params->temp > 0.0f ? params->temp : hparams.gen_flow_temp;
// Config.default_temperature, for a caller that does not set one
const float temp = params->temp > 0.0f ? params->temp : 0.7f;
std::normal_distribution<float> dist(0.0f, std::sqrt(temp));
std::vector<float> noise(ggml_nelements(t));
for (auto & v : noise) {
tools/mtmd/mtmd.cpp
+4 -2
View File
@@ -1834,7 +1834,6 @@ mtmd_gen_inp mtmd_gen_inp_default(const mtmd_context * ctx) {
return inp;
}
const clip_hparams * hparams = clip_get_hparams(ctx->ctx_gen_a);
switch (clip_get_projector_type(ctx->ctx_gen_a)) {
case PROJECTOR_TYPE_QWEN3TTS_GEN:
// https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/blob/main/generation_config.json
@@ -1843,7 +1842,10 @@ mtmd_gen_inp mtmd_gen_inp_default(const mtmd_context * ctx) {
inp.temp = 0.9f; // TODO: handle this on graph
break;
case PROJECTOR_TYPE_POCKETTTS_GEN:
inp.temp = hparams->gen_flow_temp;
// https://github.com/kyutai-labs/pocket-tts/blob/main/pocket_tts/default_parameters.py
inp.top_k = 50;
inp.top_p = 1.0f;
inp.temp = 0.7f;
break;
default:
break;