trunk/841e3dc165963ca6550560eab70a84d1c166b9eb: [FSDP2] Fix resharding to a smaller mesh after forward (#200985)
- PyTorch: 2410 events in the last 90 days
- PyTorch: 2378th Release in the last 90 days
- Previous: earlier the same day · trunk/79ae8043a9e9029f67a3b2e7f7517aeb194c9f0e
What happened
BC-breaking? No: configurations that worked on main give the same results, and those that failed now work or raise a named NotImplementedError . Summary (human written only) Agent details (if applicable) AI-assisted implementation and summary prepared with Claude Code, split out of #200034 at the author's request. With reshard_after_forward=int , FSDP reshards parameters after forward to flat chunks of their unshard…
Summary assembled by rule from the sources below