llama.cpp 修复 Metal mul_mm_id 激活超 f16 范围导致的 NaN
原标题：b10994
AI 摘要
llama.cpp 修复了 Metal 后端 mul_mm_id 算子在激活值超出 f16 范围（65504）时产生 NaN 的问题。原因是 simdgroup MMA 将 src1 窄化为 half 导致 inf 并污染整个 8x8 累加器，修复方式是按 2 的幂对 src1 重新缩放并在存储时还原，保证结果精确且对现有模型比特级一致。该缺陷曾导致 Mistral Small 4 在 Metal 上 prefill 超过 32 token 时输出全 NaN，修复后已恢复正常，性能开销约 1.14% 中位数。
正文节选
<details open> metal: fix NaN in mul_mm_id when activations exceed f16 range (#26223) * test-backend-ops: reproduce MUL_MAT_ID NaN for activations beyond f16 The Metal mul_mm_id path narrows src1 to `half` for the simdgroup MMA (`S1 = half` in every instantiation; ggml-metal.metal:10582 and :10595, mirrored at :10643/:10654 in the tensor-ops path). f16 saturates at 65504, so a model whose activations exceed that produces inf, and `simdgroup_multiply_accumulate` then turns the whole 8x8 accumu