PyTorch sparse memory bug and NCCL key leak

A PyTorch report describes silent wrong writes and heap corruption in sparse in-place addition, and an automated review flags a new NCCL encryption key landing in traces. Separately, a transformers pull request to add Omnilingual ASR models is still working through routine checks.

Sparse in-place add can write the wrong memory

An opening report on the PyTorch tracker says an in-place add of a sparse COO tensor onto a dense view can write to the wrong offset or corrupt the heap. The failure is on the CPU path when the destination is a strided view, and it can also produce wrong results without an obvious error. Anyone relying on sparse updates into non-contiguous dense storage should treat those results as untrustworthy until the bug is fixed.

New NCCL encryption key flagged in traces

An automated review of a PyTorch change flags plaintext leakage of a new NCCL encryption key into traces and caches. The note is a single automated finding, not a published exploit write-up. Operators turning on that encryption path should check whether keys can end up in logs or cached artifacts.

Omnilingual ASR models proposed for transformers

A pull request to Hugging Face transformers would add Omnilingual ASR models. The discussion so far is routine CI, lint, and test failures, with one maintainer comment on architecture. The models are not a landed feature yet.