llama.cpp Release b11093 Fixes Metal Flash Attention Mask Bounds
Open-source LLM inference engine llama.cpp has released build b11093, which includes a bug fix for mask bounds in the block pre-pass of the Metal Flash Attention implementation. This routine update improves computation correctness and stability when running transformer models on Apple Silicon devices utilizing Metal acceleration. The release comes with updated pre-compiled binaries across multiple platforms including macOS, Linux, Windows, and Android, covering backends such as CUDA, Vulkan, ROCm, and SYCL.
## BACKGROUND
llama.cpp is a widely used C/C++ framework for running LLMs efficiently on consumer hardware. Flash Attention is an optimized attention algorithm that reduces memory usage and speeds up inference by processing score matrices in smaller blocks. Metal is Apple's graphics and GPU compute API designed for high performance on Apple hardware.