The ggml-org team’s llama.cpp b11866 release removes a power-of-two rescaling step from a Metal matrix-multiplication kernel used for mixture-of-experts models. The project says per expert-token scaling in the compute graph now keeps down-projection activations within the range supported by f16 before the operation runs.
That change also removes two-stage reduction kernels, their scratch buffer and an f32 dispatch path, along with tests for inputs that exceeded the f16 range. The release lists builds for Apple Silicon macOS, as well as many other platforms; it reports no speed or accuracy measurements.
Why it matters
For llama.cpp users running mixture-of-experts models on Metal, the kernel now relies on scaling performed earlier in the compute graph; the release does not say whether this changes speed or output quality.
