llama.cpp Release b10769 Fixes Apple Metal Low-Memory Query Bug
llama.cpp version b10769 was released, featuring a targeted bug fix for Apple Metal memory querying under low-memory conditions. The update simplifies and streamlines the code responsible for checking available GPU memory on Apple devices. This release ensures greater stability for users running large language models on Mac and iOS hardware operating near memory limits. Accurate memory checks prevent runtime errors and improper resource allocation during memory-constrained inference. Introduced in PR #27701 with contributions from Niklas Wenzel and Georgi Gerganov, the fix refactors the Metal memory query routine into a more concise form. Pre-built binaries are provided across multiple platforms including Apple Silicon, Windows CUDA/Vulkan, and Linux.
## BACKGROUND
llama.cpp is a high-performance open-source LLM inference framework written in C/C++ that enables local AI deployment on consumer hardware. Apple Metal is a low-level API designed by Apple to give software direct hardware-accelerated compute access to GPUs across macOS and iOS devices.