b8670
📦 llama-cppView on GitHub →
✨ 6 features🐛 2 fixes🔧 6 symbols
Summary
This release introduces comprehensive support for HunyuanOCR models, including vision capabilities and a new chat template. It also includes various fixes related to token IDs and tensor mappings during conversion.
✨ New Features
- Added support for HunyuanOCR, including text and vision models.
- Implemented HunyuanOCR vision projector (perceiver-based) with Conv2d merge.
- Added separate HUNYUAN_OCR chat template using content-before-role format.
- Added support for xdrope RoPE scaling type.
- Added tensor mappings for perceiver projector (mm.before_rms, mm.after_rms, etc.).
- Registered HunYuanVLForConditionalGeneration for both text and mmproj conversion.
🐛 Bug Fixes
- Fixed handling of HunyuanOCR's invalid pad_token_id=-1 in converter.
- Fixed EOS/EOT token IDs retrieved from generation_config.json.