llama.cpp makes Metal fail closed on matmul shapes with no F16 kernel instead of crashing on a nil pipeline

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10691, published 2026-08-30, fixes a null-pipeline crash for F16 src1 mul_mat and mul_mat_id on Metal by declining unsupported shapes in supports_op and aborting on a nil pipeline in encoder_set_pipeline, sharing the mm dispatch between the two paths so support checks and execution cannot disagree. Build b10690 the same morning fixes a separate crash where a context shift on an unquantized K cache copied a Hadamard matrix to a k_rot tensor with no buffer assigned. Both are Apple-silicon crash fixes for anyone running long-context local inference.