TL;DR: llama.cpp addressed a critical race condition in its Metal backend's Mixture-of-Experts (MoE) reduce fusion kernel, preventing non-deterministic wrong results.
Summary: A fix was implemented in llama.cpp (commit b11735) for a threadgroup-vs-threadgroup race condition in the Metal backend's fused moe_reduce kernel. This race occurred when the destination tensor's memory overlapped with its input tensors, leading to rare, non-deterministic incorrect results. The solution rejects fusion when such an overlap is detected, ensuring correctness.
Why it matters: This fix improves the reliability and accuracy of MoE model inference on Apple Metal hardware for llama.cpp users. Developers should update their llama.cpp installations to leverage this stability improvement for local LLM deployments.
Source: github_releases