b9664
📦 llama-cppView on GitHub →
✨ 3 features🐛 1 fixes🔧 1 symbols
Summary
This release enhances SYCL backend support for reordered quantization formats (Q4_K, Q5_K, Q6_K) in MoE operations and improves fallback behavior for unsupported 3D reorder cases. Various pre-built binaries for different platforms are provided.
✨ New Features
- SYCL backend now supports reordered Q4_K and Q5_K MoE MUL_MAT_ID operations.
- SYCL backend extends MoE reorder support to Q6_K mul_mat_id.
- SYCL backend adds Q5_K reordered DMMV coverage.
🐛 Bug Fixes
- Unsupported 3D reorder cases in SYCL MoE operations now fall back gracefully instead of aborting.