Will Cromar
Will Cromar
It would be great if we could tag failures with causes such as infrastructure flakiness, model regressions, or erroneous test configs. It may also be useful to link to bugs...
Fixes bug where these functions didn't accept `List[int]` (only `int` positional args). Also enable their opinfo tests. Removed corresponding tests in `test_functions` to de-duplicate with `test_ops`. Removed some unused imports....
- Move collective op implementations out of `ProcessGroup` and into registered `__torch_dispatch__` ops. - Rewrite ProcessGroup implementation using functional collectives - `torch.distributed` dispatches to `c10d.*` (e.g. `torch.ops.c10d.allreduce_`) instead of the...