Jason Ansel
Jason Ansel
> @jansel the motivation #128063. i think for fp8 they were hoping to have a working solution in the nearer term. my understanding with cooperative reductions is it wouldn't necessarily...
The other thing you could try is increasing/decreasing: https://github.com/pytorch/pytorch/blob/ecbb8a880028e59920b013f64805a883ed2958f5/torch/_inductor/runtime/triton_heuristics.py#L1766 and (set them to be equal): https://github.com/pytorch/pytorch/blob/ecbb8a880028e59920b013f64805a883ed2958f5/torch/_inductor/runtime/hints.py#L16 That heuristic is the same ideas as the split factor in split reductions, and...
Tests are all in https://github.com/pytorch/pytorch/tree/main/test/inductor If you want to assert something about the output code you could do: https://github.com/pytorch/pytorch/blob/03cccaa76ad62a64a251587bba6d21f74b70af39/test/inductor/test_cooperative_reductions.py#L27-L37
@pytorchbot merge
I approved CI
Why is this PR making tests fail? Is the assign actually redundant in those cases?
Sounds good to me.
@pytorchbot merge
@pytorchbot merge
@galexite that test failure looks real