PyTorch MKLDNN zero_() heap overflow on half precision

A heap buffer overflow in PyTorch's MKLDNN CPU path for zero_() on bfloat16 and float16 tensors is the main item. A separate proposal for a thread-local deterministic algorithms scope has drawn almost no review.

MKLDNN zero_() overflows on BF16 and FP16 tensors

A PyTorch report says MKLDNN CPU zero_() casts bfloat16 and float16 tensor data to float pointers, writing past the allocated buffer and causing heap corruption. The thread has two messages from two participants. The failure is a memory-safety bug on a common half-precision CPU path, so anyone running those tensors through MKLDNN has a concrete reason to track a fix.

Thread-local deterministic algorithms scope proposed

A pull request adds torch.deterministic_policy, a thread-local scope for deterministic algorithms. So far the only activity is a single bot review, with no discussion.