Jeffrey Wan

@pytorchbot merge -f "Unrelated failure"

> Given those two points, it seems likely that when autograd spawns threads, we aren't properly copying over the threadlocal AMP state from the spawned thread? Yup - today the...

@zqigolden Thanks for the update! I added a comment on testing + a couple more nits, but looks ok overall. Sorry for the back and forth, but I think we...

@zqigolden Thanks for the update, this looks good to me! We just need to fix lint and it should be good to go.

@pytorchbot merge

@pytorchbot merge

@yhl48 derivative formulas can be found in: https://github.com/pytorch/pytorch/blob/62eb7a2e97205cad35ae7acd02a6bbf835db341f/tools/autograd/derivatives.yaml#L623-L626 Also see some background around complex gradients: https://pytorch.org/docs/stable/notes/autograd.html#autograd-for-complex-numbers

@yhl48 To actually skip the test so that CI doesn't fail, you can add something like: ``` DecorateInfo(unittest.expectedFailure, 'TestUnaryUfuncs', 'test_reference_numerics_large', device_type='cuda', dtypes=[torch.cdouble]), ``` (similar to what the OpInfo for `exp`...

@pytorchbot merge

Sorry we typically don't accept these types of PRs. If you would like to contribute I'd look at issues tagged "good first issue".