trunk/4310195f8a7c4a8bc462b56a073c70fc9c9bbf64: [c10d] Release unused functional Work results (#193239)
- PyTorch: 138 events in the last 90 days
- PyTorch: 137th Release in the last 90 days
- Previous: earlier the same day · trunk/5818c805bb5308e0fb8634f3f07196b172ba410a: Create a @periodic Decorator (#191946)
What happened
NCCL2 queues caller-visible Work objects for watchdog and CUDA graph tracking. Those references retain collective tensors after wait_tensor removes the functional registry entry. Queue only shared completion state and an input tensor shelf, and retain only completion state during graph capture. Caller waits clear the shared shelf, including every child of a coalesced operation, while direct Work results remain avail…
Summary assembled by rule from the sources below