viable/strict/1790621778: Improve async collective remap error (#195691)
- PyTorch: 1594 events in the last 90 days
- PyTorch: 1575th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790620535: [c10d] Allow CPU device_id in init_process_group (#198340)
What happened
Fixes #119890 Updates the Dynamo remap failure for legacy torch.distributed collectives called with async_op=True so the user-facing message explains: only synchronous async_op=False legacy collectives are remapped automatically; users who do not need a Work handle can set async_op=False ; users who do need async semantics should rewrite with torch.distributed._functional_collectives and synchronize via wait_tensor…
Summary assembled by rule from the sources below