b9933
📦 llama-cppView on GitHub →
🐛 3 fixes🔧 3 symbols
Summary
This release focuses on stability and correctness for OpenCL operations, specifically fixing Q6_K GEMM/GEMV issues on Adreno GPUs when weight dimensions are not multiples of 128, and resolving potential buffer corruption via improved alignment handling.
🐛 Bug Fixes
- Fixed garbled output for Q6_K weights where ne01 is not a multiple of 128 on Adreno GPUs.
- Implemented alignment slack reservation in OpenCL buffer allocation to prevent subbuffer overlap corruption when carving quantized weights.
- Routed Q6_K dense matrix multiplication (GEMM/GEMV) where ne01 is not a multiple of 128 off the noshuffle path, falling back to CPU for the GEMM kernel in those specific cases.