← Back to events
ActiveAIRelease

trunk/be904b2055a51551ceb3cc272b4d772bac22a470: Make sparse tensor validation state thread-local (#194751)

Photo: PyTorch Releases

What happened

Fixing an Issue Issue Fixes #194750 Summary QED Audit is proposing the thread-local fix requested by @malfet for #194750 . AI-assisted content (Codex): Move the deferred sparse-validation collector onto PyTorch's existing _SerializationLocal state so concurrent torch.load calls cannot validate or clear one another's tensors. Add a deterministic integration test that pauses a malformed weights-only load at validation…

Summary assembled by rule from the sources below

Why it's spreading

Sources