PyTorch dynamo fixes and WeatherNext 2

PyTorch drew most of the day's AI and ML attention with correctness, performance, and security reports, while Hugging Face Transformers gained a weather forecasting model. Documentation and lower-priority feature requests filled out a quiet technical slate.

Dynamo drops mutations on local-class methods

A contributor reported that mutations of a closure-captured list or dict are silently dropped when they happen inside a method of a locally defined class instantiated in a torch.dynamo compiled function under fullgraph=True with no graph break. The report includes a root-cause analysis and a proposed fix. Silent data loss without a graph break matters for anyone relying on fullgraph compilation for correctness.

Parallel CPU Col2im for lower memory traffic

A pull request proposes processing Col2im channels in parallel on CPU to reduce memory access. The author reports roughly 10x speedups on specific fold workloads. The change targets remaining CPU paths in convolution-related transforms where memory traffic still dominates.

WeatherNext 2 lands in Transformers

A pull request adds the WeatherNext 2 weather forecasting model to Hugging Face Transformers. Review comments focused on feature extractors, casts, and repeated CI runs. The addition expands pretrained options for atmospheric prediction inside the library.

safe_globals can widen temporary allowlist entries

A report describes an inconsistency in PyTorch's safe_globals context manager that can promote temporary entries into a process-wide allowlist. The behavior is security-relevant for code that expects only scoped serialization allowlisting. Users of torch load safeguards should treat the manager as weaker than documented until fixed.

CUDA RNG output tied to SM count

A documentation pull request notes that CUDA generator output depends on the device SM count and proposes updates to the reproducibility note and seeding docstrings. Discussion questioned whether bit-exact CUDA RNG can be guaranteed across GPUs with different SM counts. The clarification affects anyone chasing cross-device reproducibility.