llama.cpp Release b10691 Fixes Metal Pipeline Crash on Apple Silicon
Release b10691 of llama.cpp addresses a null-pipeline crash occurring during F16 matrix multiplication operations on Apple's Metal framework. The update introduces a fail-closed mechanism for shapes missing F16 kernels and aborts execution on nil pipelines. This fix prevents application crashes for macOS and iOS users running local LLM inference using F16 precision on Apple Silicon GPUs. It ensures more stable performance and reliability when executing matrix operations on Apple hardware. The release implements an abort mechanism in `encoder_set_pipeline` when a nil pipeline is detected and shares the matrix multiplication dispatch logic with `supports_op`. Additionally, the KleidiAI-enabled macOS Apple Silicon build has been disabled in this release.
## BACKGROUND
llama.cpp is a popular open-source C/C++ library designed for efficient local inference of large language models like Meta's Llama. Metal is Apple's low-level graphics and compute API that allows developers to leverage Apple Silicon GPUs for accelerated workloads. Matrix multiplication, particularly using F16 precision, is a fundamental mathematical computation used during LLM inference.