Skip to content

[Loss] Preserve caller FSDP policies across chunked backward - #5563

Open
ZenAlexa wants to merge 13 commits into
pytorch:mainfrom
ZenAlexa:contrib/5408-chunked-loss-deferred-sync
Open

ZenAlexa wants to merge 13 commits into
pytorch:mainfrom
ZenAlexa:contrib/5408-chunked-loss-deferred-sync