Fix several test_ops cuda dtypes tests #60922
Conversation
💊 CI failures summary and remediationsAs of commit 05fca2e (more details on the Dr. CI page and at hud.pytorch.org/pr/60922): ✅ None of the CI failures appear to be your fault 💚
🚧 1 fixed upstream failure:These were probably caused by upstream breakages that were already fixed.
Please rebase on the
|
jbschlosser
added
the
triaged
| @@ -6038,7 +6038,8 @@ def gradcheck_wrapper_triangular_input(op, input, *args, upper=False, **kwargs): | |||
| dtypesIfCPU=all_types_and_complex(), | |||
| dtypesIfCUDA=floating_and_complex_types_and(torch.float16, *[torch.bfloat16] if CUDA11OrLater else []), | |||
| dtypesIfROCM=floating_types_and(torch.half, torch.bfloat16), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.float16), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.float16, | |||
| *[torch.bfloat16] if SM60OrLater else []), | |||
There was a problem hiding this comment.
I don't think the sm_60 check would be sufficient, as bfloat16 should have been introduced in CUDA11 (https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html#cuda-general-new-features), so this could fail on 10 and previous releases. Wouldn't if CUDA11OrLater as done for the forward pass also work here?
There was a problem hiding this comment.
It unfortunately failed in my previous commit which was using CUDA11OrLater, see https://app.circleci.com/pipelines/github/pytorch/pytorch/343601/workflows/695dad4b-86d7-49f4-b4c5-e2aa7b1d0bfa/jobs/14465553
There was a problem hiding this comment.
We may use something like (CUDA11OrLater and SM60OrLater).
There was a problem hiding this comment.
Did you want to try that change, @xwang233?
There was a problem hiding this comment.
The skip on line 6079 should be removed once the dtypes are corrected
|
Added the ci/master label so this tests against more configs. I also suspect this is a combination of CUDA11OrLater + a particular SM. |
| @@ -6545,7 +6546,8 @@ def gradcheck_wrapper_triangular_input(op, input, *args, upper=False, **kwargs): | |||
| dtypesIfCPU=all_types_and_complex(), | |||
| dtypesIfCUDA=floating_types_and(torch.float16, *[torch.bfloat16] if CUDA11OrLater else [], | |||
| torch.complex64, torch.complex128), | |||
| backward_dtypesIfCUDA=floating_types_and(torch.float16, torch.complex64, torch.complex128), | |||
| backward_dtypesIfCUDA=floating_types_and(torch.float16, *[torch.bfloat16] if SM60OrLater else [], | |||
There was a problem hiding this comment.
The skip on line 6557 needs to be removed
| @@ -6904,7 +6906,7 @@ def gradcheck_wrapper_triangular_input(op, input, *args, upper=False, **kwargs): | |||
| op=lambda tensors, equation: torch.einsum(equation, tensors), | |||
| dtypes=all_types_and_complex_and(torch.half, torch.bfloat16), | |||
| dtypesIfCUDA=floating_and_complex_types_and(torch.half, *[torch.bfloat16] if CUDA11OrLater else []), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.half), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.half, *[torch.bfloat16] if SM60OrLater else []), | |||
There was a problem hiding this comment.
The skip on line 6915 needs to be removed
| @@ -6069,13 +6069,11 @@ def gradcheck_wrapper_triangular_input(op, input, *args, upper=False, **kwargs): | |||
| dtypesIfCPU=all_types_and_complex(), | |||
| dtypesIfCUDA=floating_and_complex_types_and(torch.float16, *[torch.bfloat16] if CUDA11OrLater else []), | |||
| dtypesIfROCM=floating_types_and(torch.half, torch.bfloat16), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.float16), | |||
| backward_dtypesIfCUDA=floating_and_complex_types_and(torch.float16, | |||
| *[torch.bfloat16] if (SM60OrLater and CUDA11OrLater) else []), | |||
There was a problem hiding this comment.
ping @mruberry . Seems like the CI is happy with the new flags
|
@mruberry has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Close #60443