← Back to events
ActiveAIRelease

viable/strict/1790789324: Fix _int_mm CPU corner case (#194682)

Photo: PyTorch Releases

What happened

This is to fix #195066 _mkldnn_gemm_i8i8i32_with_blas took leading dimensions from the input strides, but the stride of a size-1 dim is arbitrary and inductor emits 0 for it in some cases. On the other hand, oneDNN's row-major gemm requires lda >= K so it will silently reject the call, causing accuracy failures in workloads like int8 quantized LLM decoding. The PR fixes this corner case and also check this oneDNN ge…

Summary assembled by rule from the sources below

Why it's spreading

Sources