enne2/qwentts.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cleaning

Browse Source
This commit is contained in:
Pascal
2026-06-21 17:30:21 +02:00
parent 26fcea5468
commit 4536dcdce2
2 changed files with 19 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/code-predictor-weights.h
+9 -11
View File
@@ -42,7 +42,6 @@ struct CodePredictorWeights {
int num_key_value_heads; int num_key_value_heads;
int head_dim; int head_dim;
int vocab_size; int vocab_size;
int max_position_embeddings;
int num_acoustic_codebooks; // num_code_groups - 1 int num_acoustic_codebooks; // num_code_groups - 1
float rope_theta; float rope_theta;
float rms_norm_eps; float rms_norm_eps;
@@ -68,16 +67,15 @@ struct CodePredictorWeights {
}; };
static bool code_predictor_weights_load(CodePredictorWeights * cw, const GGUFModel & gf, ggml_backend_t backend) { static bool code_predictor_weights_load(CodePredictorWeights * cw, const GGUFModel & gf, ggml_backend_t backend) {
cw->hidden_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.embedding_length"); cw->hidden_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.embedding_length");
cw->intermediate_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.feed_forward_length"); cw->intermediate_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.feed_forward_length");
cw->num_hidden_layers = (int) gf_get_u32(gf, "qwen3-tts.code_pred.block_count"); cw->num_hidden_layers = (int) gf_get_u32(gf, "qwen3-tts.code_pred.block_count");
cw->num_attention_heads = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.head_count"); cw->num_attention_heads = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.head_count");
cw->num_key_value_heads = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.head_count_kv"); cw->num_key_value_heads = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.head_count_kv");
cw->head_dim = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.key_length"); cw->head_dim = (int) gf_get_u32(gf, "qwen3-tts.code_pred.attention.key_length");
cw->vocab_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.vocab_size"); cw->vocab_size = (int) gf_get_u32(gf, "qwen3-tts.code_pred.vocab_size");
cw->max_position_embeddings = (int) gf_get_u32(gf, "qwen3-tts.code_pred.context_length"); cw->rope_theta = gf_get_f32(gf, "qwen3-tts.code_pred.rope.freq_base");
cw->rope_theta = gf_get_f32(gf, "qwen3-tts.code_pred.rope.freq_base"); cw->rms_norm_eps = gf_get_f32(gf, "qwen3-tts.code_pred.attention.layer_norm_rms_epsilon");
cw->rms_norm_eps = gf_get_f32(gf, "qwen3-tts.code_pred.attention.layer_norm_rms_epsilon");
int num_code_groups = (int) gf_get_u32(gf, "qwen3-tts.num_code_groups"); int num_code_groups = (int) gf_get_u32(gf, "qwen3-tts.num_code_groups");
if (num_code_groups <= 1) { if (num_code_groups <= 1) {
src/talker-weights.h
+10 -18
View File
@@ -73,10 +73,6 @@ struct TalkerWeights {
int num_key_value_heads; int num_key_value_heads;
int head_dim; int head_dim;
int vocab_size; int vocab_size;
int text_vocab_size;
int text_hidden_size;
int max_position_embeddings;
int position_id_per_seconds;
float rope_theta; float rope_theta;
float rms_norm_eps; float rms_norm_eps;
int mrope_section_t; int mrope_section_t;
@@ -100,20 +96,16 @@ struct TalkerWeights {
}; };
static bool talker_weights_load(TalkerWeights * tw, const GGUFModel & gf, ggml_backend_t backend) { static bool talker_weights_load(TalkerWeights * tw, const GGUFModel & gf, ggml_backend_t backend) {
tw->hidden_size = (int) gf_get_u32(gf, "qwen3-tts.talker.embedding_length"); tw->hidden_size = (int) gf_get_u32(gf, "qwen3-tts.talker.embedding_length");
tw->intermediate_size = (int) gf_get_u32(gf, "qwen3-tts.talker.feed_forward_length"); tw->intermediate_size = (int) gf_get_u32(gf, "qwen3-tts.talker.feed_forward_length");
tw->num_hidden_layers = (int) gf_get_u32(gf, "qwen3-tts.talker.block_count"); tw->num_hidden_layers = (int) gf_get_u32(gf, "qwen3-tts.talker.block_count");
tw->num_attention_heads = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.head_count"); tw->num_attention_heads = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.head_count");
tw->num_key_value_heads = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.head_count_kv"); tw->num_key_value_heads = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.head_count_kv");
tw->head_dim = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.key_length"); tw->head_dim = (int) gf_get_u32(gf, "qwen3-tts.talker.attention.key_length");
tw->vocab_size = (int) gf_get_u32(gf, "qwen3-tts.talker.vocab_size"); tw->vocab_size = (int) gf_get_u32(gf, "qwen3-tts.talker.vocab_size");
tw->text_vocab_size = (int) gf_get_u32(gf, "qwen3-tts.talker.text_vocab_size"); tw->rope_theta = gf_get_f32(gf, "qwen3-tts.talker.rope.freq_base");
tw->text_hidden_size = (int) gf_get_u32(gf, "qwen3-tts.talker.text_hidden_size"); tw->rms_norm_eps = gf_get_f32(gf, "qwen3-tts.talker.attention.layer_norm_rms_epsilon");
tw->max_position_embeddings = (int) gf_get_u32(gf, "qwen3-tts.talker.context_length"); tw->mrope_interleaved = gf_get_bool(gf, "qwen3-tts.talker.rope.mrope_interleaved");
tw->position_id_per_seconds = (int) gf_get_u32(gf, "qwen3-tts.talker.position_id_per_seconds");
tw->rope_theta = gf_get_f32(gf, "qwen3-tts.talker.rope.freq_base");
tw->rms_norm_eps = gf_get_f32(gf, "qwen3-tts.talker.attention.layer_norm_rms_epsilon");
tw->mrope_interleaved = gf_get_bool(gf, "qwen3-tts.talker.rope.mrope_interleaved");
std::vector<uint32_t> mrope = gf_get_array_u32(gf, "qwen3-tts.talker.rope.mrope_section"); std::vector<uint32_t> mrope = gf_get_array_u32(gf, "qwen3-tts.talker.rope.mrope_section");
if (mrope.size() == 3) { if (mrope.size() == 3) {