viable/strict/1792492019: [FSDP2] Separate collective copy preparation from execution (#200034)
- PyTorch: 2589 events in the last 90 days
- PyTorch: 2554th Release in the last 90 days
- Previous: earlier the same day · trunk/9544ec5e23cef8a0a4ac3f44c27e0e94d651ad02: [Testcase Refactoring] Label P2PIpcTest as CUDA (#194849)
What happened
Keep the original AG/RS preparation and copying as defaults. AI-assisted implementation and summary prepared with Codex and Claude Code for the requested change above. FSDP now prepares each collective copy separately from executing it, so the next PR can register native copies. To review, start with Note [All-gather output fn] and Note [Reduce-scatter input fn] in _fsdp_collectives.py and AllGatherInput in _fsdp_ap…
Summary assembled by rule from the sources below