b9019
Breaking Changes📦 llama-cppView on GitHub →
⚠ 1 breaking✨ 3 features🐛 4 fixes🔧 7 symbols
Summary
This release focuses heavily on internal refactoring, moving model loading utilities (`load_hparams`, `load_tensors`) to be model-specific, and introducing build graph capabilities. Numerous minor fixes and cleanups were also performed.
⚠️ Breaking Changes
- The functions `load_hparams` and `load_tensors` have been moved from a general location to be defined per-model, which may break downstream code relying on their previous location.
Migration Steps
- Update code that calls `load_hparams` and `load_tensors` to reflect their new per-model definition location.
- Remove any migration scripts or ifdef blocks related to previous model loading logic, as the migration script was removed.
✨ New Features
- Added `build_graph` functionality.
- Added `llm_arch_model_i` support.
- Added back `create_tensor_qkv`.
🐛 Bug Fixes
- Fixed issues with missing file includes.
- Enforced one `llm_build` per architecture.
- Fixed build errors.
- Fixed ordering issues.