viable/strict/1792521211: [funcol] Add broadcast autograd (#200882)
- PyTorch: 2614 events in the last 90 days
- PyTorch: 2575th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1792517296: [fx] Make _Patcher's already-patched guard actually fire (#201464)
What happened
_c10d_functional.broadcast has no autograd, unlike torch.distributed.nn.functional.broadcast . Fixes #200877 . Backward sums the grads onto src with the functional reduce, which returns zeros off src. This keeps backward rank agnostic, so it traces the same graph on every rank. Test plan: python -m pytest test/distributed/test_functional_differentials.py This PR was authored with an AI assistant (Claude). Pull Reque…
Summary assembled by rule from the sources below