b9049
📦 llama-cppView on GitHub →
✨ 2 features🐛 6 fixes🔧 6 symbols
Summary
This release introduces support for MiniCPM-V 4.6 and enables flash attention capabilities. It also includes numerous internal fixes and cleanups related to model conversion and code style.
Migration Steps
- Rename 'clip_graph_minicpmv4_6' if it was used previously.
- Update usage to use 'TYPE_MINICPMV4_6' instead of the previous type identifier.
- Update model conversion scripts to use filter_tensors and include 'model.vision_tower'.
✨ New Features
- Support for MiniCPM-V 4.6 added in a new branch.
- Enabled flash attention support by using build_attn.
🐛 Bug Fixes
- Fixed a general code bug.
- Fixed pre-commit hooks.
- Fixed conversion issues.
- Fixed chkhsh related issues.
- Fixed type checking issues.
- Fixed code style issues.