PyTorch ROCm FSDP and Linear matmul proposals

PyTorch saw two low-activity technical proposals on ROCm distributed training support and a Linear operator change. Both remain early with limited discussion.

MoRI collectives requested for FSDP on ROCm

An opening issue in the PyTorch repository asks for MoRI collectives to support FSDP on ROCm and AMD Pollara. No discussion has occurred yet. Readers tracking AMD GPU training paths may watch whether the request advances distributed features on that stack.

Linear inner_product calls switched to matmul

A pull request proposes replacing Linear inner_product calls with matmul. Automated review notes a possible repeated weight reorder concern in ideep. The change could affect how Linear layers are lowered and optimized in that backend.