b8705
📦 llama-cppView on GitHub →
✨ 6 features🐛 1 fixes🔧 3 symbols
Summary
This release introduces support for the step3-vl-10b model and includes several internal optimizations and refactoring, such as using fused QKV and updating parameter handling for MmprojModel.
✨ New Features
- Support for the step3-vl-10b model.
- Use fused QKV and mapping tensor in tensor_mapping.py.
- Guard hardcoded parameters and drop crop metadata.
- Get understand_projector_stride from global config.
- Move img_u8_resize_bilinear_to_f32 into the step3vl class.
- Added width and heads to MmprojModel.set_gguf_parameters.
🐛 Bug Fixes
- Fixed an issue related to \r\n characters (mess).