[inductor] remove fft and svd ops from fake_incorrect_kernels

[nkaretnikov]

Stack from ghstack (oldest at bottom):

• -> [inductor] remove fft and svd ops from fake_incorrect_kernels #103616

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @ngimel @yf225 @chenyang78

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/103616

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure

As of commit b5b2c93:

NEW FAILURE - The following job has failed:

• linux-bionic-cuda12.1-py3.10-gcc9-sm86 / test (slow, 2, 2, linux.g5.4xlarge.nvidia.gpu) (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[eellison]

LGTM if CI Passes

[eellison]

actually, can you remove https://github.com/pytorch/pytorch/blob/main/test/test_ops.py#L1969-L1990 as part of this ?

[nkaretnikov]

@eellison would failures in fake_tensor_stride_failing_ops be a blocker for merging this if that fails? or can I merge without?

[eellison]

Yea.. If there are striding failures then I think we shouldn't land this. Otherwise we'll get runtime assert failures in inductor

[nkaretnikov]

@eellison Please take another look: I've pushed an update implementing stride logic from _exec_fft in meta_fft_c2c.

@XiaobingSuper Could you test if this PR fixes the issues you were having with GoogleFnet? This PR now runs the repro from #88248 (comment) instead of erroring out (note: the trailing paren is missing in the repro).

Using FallbackKernel: torch.ops.prims.convert_element_type.default
Using FallbackKernel: aten._fft_c2c.default
tensor([[[  740.4529+0.0000j,   126.5869-135.1518j, -1001.8738-182.1360j,
...

[nkaretnikov]

@pytorchbot rebase

[pytorchmergebot]

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

[pytorchmergebot]

Successfully rebased gh/nkaretnikov/138/orig onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via ghstack checkout https://github.com/pytorch/pytorch/pull/103616)

[eellison]

😍 😍 😍

[peterbell10]

I believe cpu and cuda results diverge in some cases because cuFFT can only transform up to 3 dimensions at once, so we end up taking multiple passes for higher dimensions. This is still better than assuming contiguous tough.

[nkaretnikov]

@pytorchbot merge -f "the slow test failure is also on hud"

[pytorchmergebot]

Merge started

Your change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[XiaobingSuper]

@eellison Please take another look: I've pushed an update implementing stride logic from _exec_fft in meta_fft_c2c.

@XiaobingSuper Could you test if this PR fixes the issues you were having with GoogleFnet? This PR now runs the repro from #88248 (comment) instead of erroring out (note: the trailing paren is missing in the repro).

Using FallbackKernel: torch.ops.prims.convert_element_type.default
Using FallbackKernel: aten._fft_c2c.default
tensor([[[  740.4529+0.0000j,   126.5869-135.1518j, -1001.8738-182.1360j,
...

@nkaretnikov, thanks, I checked it on my side, and it works now.