A llama.cpp bug had silently disabled Metal norm/MUL fusion, costing about 5% token-generation speed
GitHub·medium signal
llama.cpp build b10909 (2026-09-11) moves all Metal fusion patterns into one table. In doing so, the author found that the fusion query passed a relative instead of absolute node index, so the last node of every pattern failed the use-count check and almost no fusions ran. The commit calls this a roughly 5% token-generation regression on Apple Silicon, now fixed. The commits carry an 'Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731' trailer.