diff options
| author | soryu <soryu@soryu.co> | 2026-01-30 03:07:52 +0000 |
|---|---|---|
| committer | soryu <soryu@soryu.co> | 2026-01-30 03:07:52 +0000 |
| commit | c526f93aa4255cb581eeb3f7a495c1689683b0a2 (patch) | |
| tree | fbdc579d04fe92dc610ec8c84b77eeffb9141622 /makima/src/llm/mod.rs | |
| parent | a9655dccdad116db2b92c13794ddd559f160148d (diff) | |
| download | soryu-c526f93aa4255cb581eeb3f7a495c1689683b0a2.tar.gz soryu-c526f93aa4255cb581eeb3f7a495c1689683b0a2.zip | |
Fix Qwen3-TTS tensor paths to match HuggingFace model structure
The HuggingFace model uses different tensor name prefixes:
- talker.model.text_embedding instead of model.embed_tokens
- talker.codec_head instead of lm_head
- talker.code_predictor.model.codec_embedding instead of code_embeddings
- talker.code_predictor.lm_head instead of output_heads
Also removed input_proj which doesn't exist in the HF model.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Diffstat (limited to 'makima/src/llm/mod.rs')
0 files changed, 0 insertions, 0 deletions