PyTorch attention and kernel security issues

PyTorch received two security-related reports on causal attention masking and compiled kernel bounds checks, with additional low-priority proposals on cuDNN eligibility and quantization. Nearly all activity stayed inside the main PyTorch repository, plus one unanswered Transformers feature request.

Causal mask leaks future tokens on MPS half precision

A PyTorch bug report states that scaled_dot_product_attention with is_causal=True on the MPS backend silently fails for float16 and bfloat16, allowing future token leakage. The failure is described as silent rather than raising an error, so causal integrity can break without obvious warning. Users running transformer workloads on Apple Metal in reduced precision have direct reason to track the issue.

Silent out-of-bounds read fixed in strided compiled kernels

A pull request fixing a silent out-of-bounds read for strided or offset compiled kernels was merged by bot in the PyTorch repository. The change closes a path that could read beyond intended memory without signaling an error. Anyone relying on PyTorch compiled kernels with non-contiguous layouts benefits from the hardened behavior.

Input-layout check proposed to restore cuDNN Flash fallback

A PyTorch issue proposes adding an input-layout eligibility check for cuDNN so that short padded queries can again fall back to FlashAttention. The intent is to avoid unnecessary rejection of valid layouts and keep the faster path available. Workloads that pad short sequences stand to regain consistent performance if the check lands.

Missing alias check flagged for scaled_addmm_ scales

A bot-raised PyTorch issue notes the absence of an alias check between scales and output in scaled_addmm_, creating a risk of cuBLASLt races. Overlapping memory could produce incorrect results in scaled matrix multiplies. CUDA users of low-precision linear layers should be aware of the potential hazard.

FP8 quantization support sought for Xiaomi multimodal model

A single feature request in Hugging Face Transformers asks for FP8 and MoE compatibility with the Xiaomi-MiMo-V2.5 multimodal models. No replies have appeared yet. Parties evaluating efficient quantization of that model family have little concrete progress to follow so far.