PyTorch hub locking and CUDA Blackwell flags

PyTorch development focused on concurrent download safety in torch.hub and CUDA architecture detection for Blackwell GPUs. A separate low-activity request sought finer CPU offload controls in llama.cpp.

PyTorch hub serializes cold-cache downloads

A pull request in pytorch/pytorch adds FileLock around cold-cache downloads in torch.hub so concurrent callers are serialized. Automated review notes NFS polling cost and a missing lower bound on the filelock dependency. The change matters for users who share hub caches across processes and risk races during first-time model fetches.

CUDA arch detection drops accelerated suffixes

An opening report in pytorch/pytorch states that auto-detected CUDA architecture flags omit accelerated suffixes, mapping sm_120a to sm_120 and breaking Blackwell NVFP4 and block-scaled MMA support. The post proposes restoring the full suffix in the detection path. Correct flags are required for developers targeting those Blackwell matrix features.

llama.cpp seeks FFN band selection

A feature request in ggml-org/llama.cpp asks to extend --n-cpu-ffn with --n-cpu-mode style band selection for automated FFN offloading. Discussion so far is limited to a few messages from two participants. The option would give operators tighter control when splitting feed-forward layers between CPU and accelerator.