llama.cpp Release b10743 Introduces Metal Optimizations for Apple M2 Pro
llama.cpp has released build b10743, introducing Metal flash-attention vector tuning optimizations specifically for Apple M2 Pro processors across new data types. The update also refreshes prebuilt binaries across supported platforms including macOS, Linux, Windows, and Android. This routine maintenance release brings targeted inference optimizations for users running local LLMs on M2 Pro Apple Silicon Macs. It highlights the project's continuous fine-tuning of hardware-accelerated kernels across specific GPU architectures. Pull Request #28122 adds dedicated flash-attention vector tuning parameters and new data type support for M2 Pro GPUs within the Metal backend. Updated release binaries were built for macOS (arm64, x64), Linux (CPU, Vulkan, ROCm, SYCL, OpenVINO), Windows (CUDA 12/13, Vulkan, ROCm), and Android.
## BACKGROUND
llama.cpp is a widely used open-source C/C++ framework for running large language models locally with high efficiency on consumer hardware. On macOS devices, it uses Apple's Metal API to execute neural network computations on Apple Silicon GPUs. FlashAttention is an algorithm designed to speed up transformer attention layers and reduce memory usage by optimizing reads and writes between GPU high-bandwidth memory and on-chip SRAM.