PyTorch fixes integer overflow in CPU bf16/fp16 GEMM staging buffer

PyTorch, the open-source deep learning framework maintained by Meta, has released a critical fix addressing an integer overflow in the BLAS GEMM (General Matrix Multiply) staging buffer size calculation. The issue specifically affected operations using bfloat16 (bf16) and float16 (fp16) precision on CPU. When performing matrix multiplications essential for neural network inference and training, the overflow caused the staging buffer to be allocated with an incorrect size — potentially too small — leading to memory corruption and incorrect numerical results.

The fix, tagged as 'viable/strict/1785041792' and committed by contributor Diacod-I on July 26, corrects the size computation to prevent integer wraparound. While the bug likely only manifested under specific conditions (large matrices or particular dimensions), its silent nature made it particularly dangerous: models could produce subtly wrong outputs without obvious errors. This patch is especially relevant for developers running PyTorch on CPU for edge deployment or environments without GPU acceleration, where bf16/fp16 are used to reduce memory and improve throughput.