Make determinism of channels_last more conservative

[kimishpatel]

Summary:
When in doubt because both is_contiguous and is_contiguous(channels_last)
return true, assume it is channels first

Differential Revision: D83998877

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14862

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit f0abe93 with merge base 8ac6300 ():

NEW FAILURE - The following job has failed:

• pull / test-samsung-models-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 385db36706cd7366228db730aaad0d8e84b0489cb03d668c8a39f3de622a43db /exec failed with exit code 1

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

[meta-codesync]

@kimishpatel has exported this pull request. If you are a Meta employee, you can view the originating Diff in D83998877.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[digantdesai]

Should we just use Tensor.dim_order(ambiguity_check=True)?

[kimishpatel]

it raises runtime error when ambiguity_check=True. Thats why i had to do this

[digantdesai]

IIUC this is about ambiguous cases right? If yes, when we have ambiguous shape [1,1,1,1] this may cause issues when the shape changes at runtime and our assumptions about the memory format (one way or the other) breaks down and results in silently calculating incorrect results.

The right solution might be to get rid of torch.memory_format from XNNPACK switch to unambiguous (at least with the new API with check) and deal with it properly.

[kimishpatel]

IIUC this is about ambiguous cases right? If yes, when we have ambiguous shape [1,1,1,1] this may cause issues when the shape changes at runtime and our assumptions about the memory format (one way or the other) breaks down and results in silently calculating incorrect results.

How would you determine the unambiguous cases? What I found for one of the models, which was not using channels in the conventional sense, it was just a 4D tensor, it resulted in failures to lower.

The right solution might be to get rid of torch.memory_format from XNNPACK switch to unambiguous (at least with the new API with check) and deal with it properly.

Yes, but that challenge is what is unambiguous. Problem is we dont have a language in pytorch to unambiguously say what is the dim order. I think the right solution would be that. Although when I think about it, it feels fairly non-trivial. User would have to be explicit in their intent within the model to tell you what is the dim order.

I think safe thing to do would be start out wit making everything contiguous/channels first and then introduce any dim order changes explicitly and this would have to happen within XNNPACK backend passes.

[GregoryComer]

IIUC this is about ambiguous cases right? If yes, when we have ambiguous shape [1,1,1,1] this may cause issues when the shape changes at runtime and our assumptions about the memory format (one way or the other) breaks down and results in silently calculating incorrect results.

The right solution might be to get rid of torch.memory_format from XNNPACK switch to unambiguous (at least with the new API with check) and deal with it properly.

Are there any specific examples of this that you're aware of that yield silent correctness issues?

[digantdesai]

How would you determine the unambiguous cases?

if a tensor of a given shape can be mapped to multiple dim_order ==> it is ambiguous.

https://github.com/pytorch/pytorch/blob/85801126821d4f509f3cf5aafa24dbcd3cd11183/torch/_tensor.py#L1532