fix(utils): Make torch_compilable_check compatible with torch.export strict mode

[harshaljanjani]

What does this PR do?

The following issue was identified and fixed in this PR:

→ Reasoning: The impact of this fix goes beyond Mask2Former and DeformableDetr and should fix any model that uses torch_compilable_check. Most users run models in eager where the fn works just fine; only torch._check_with only breaks during export tracing. Mask2Former already has test coverage for torch.export.export() functionality and it currently fails (check CI for failing test_modeling_mask2former.py::Mask2FormerModelIntegrationTest::test_export), this should fix that.
→ I decided to pick one other pattern to prove that this fix generalizes to other models without regressions using torch_compilable_check and add test coverage for it. Picked DeformableDetr because it uses the same deformable attn pattern as in Mask2Former (same torch_compilable_check() call in DeformableDetrMultiscaleDeformableAttention as in Mask2FormerPixelDecoderEncoderMultiscaleDeformableAttention) and didn't have coverage for an export test yet. Again for the test as well, I followed the canonical pattern from Mask2FormerModelIntegrationTest::test_export.
→ For more details on reproducing the bug and the output screenshots, please visit the linked issue!

Fixes #44265.

Added DeformableDetr test before the fix (feel free to cross-check; this error is reproducible):

Added DeformableDetr test after the fix (feel free to cross-check):

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you fix any necessary existing tests?
• Did you write any new necessary tests?

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deformable_detr

[ArthurZucker]

thanks a lot for fixing!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[harshaljanjani]

Thank you so much for your time @ArthurZucker! I was wondering if I may ask you to check in on this PR as it has been pending review for some time now; at your convenience and sorry if this is a bother :)

[IlyasMoutawwakil]

Hi ! the test you mention in the PR uses strict mode which we don't wanna support in general because it's too constraining, non-strict mode is generally enough for model export (even onnx export doesn't require strict mode).
Do you have any specific use case for it or was the intent to just fix the failing test ?

[harshaljanjani]

Good day @IlyasMoutawwakil!
Since Mask2FormerModelIntegrationTest::test_export used the pattern, I read it as intentional; did some RCA, noticed a broader issue, and posted the fix. The merged fix itself should be harmless either way; the only clarification I'd ask for is if it's aligned with the project's stance. Since the project stance is to not support strict mode as you mentioned, I'm happy to open a follow-up PR to switch both the existing Mask2Former test and the new DeformableDetr test to strict=False, but I'm missing a reason to revert the change I made to import_utils.py which preserves the original semantics in eager and non-strict export, while also being compatible with strict export.
Would love to know your thoughts :)

[IlyasMoutawwakil]

unfortunately the change actually does break a couple models 😅, we have a more general test for torch.export (it's marked as slow so it doesn't appear on fast ci) and some models started failing after this change because we also use torch._check to give dynamo compiler hints, basically telling dynamo that it should expect a condition to be true at compile time instead of evaluating it. I document that in the function's docstring but it might be easy to miss there.

the failing tests can be seen in here for example https://github.com/huggingface/transformers/actions/runs/22426245229/job/64935308903
i'm reverting that line and cleaning up the strict mode exports in #44293

[IlyasMoutawwakil]

to sum it up:

• torch._check is the most compatible way with torch.compile and torch.export non-strict mode
• torch._check seems to fail with torch.export strict mode (not sure how frequently).

I would argue that this requires fixing in pytorch because the torch._check api is what torch.export suggests using when you get data dependency failures, so it's weird for it to not be compatible with strict mode.

see this traceback from a failing export due to datat dep:

FAILED tests/models/vits/test_modeling_vits.py::VitsModelTest::test_torch_export - torch.fx.experimental.symbolic_shapes.GuardOnDataDependentSymNode: Could not guard on data-dependent expression u4 < 1 (unhinted: u4 < 1).  (Size-like symbols: none)

consider using data-dependent friendly APIs such as guard_or_false, guard_or_true and statically_known_true.
Caused by: (_subclasses/fake_impls.py:1135 in conv)
For more information, run with TORCH_LOGS="dynamic"
For extended logs when we create symbols, also add TORCHDYNAMO_EXTENDED_DEBUG_CREATE_SYMBOL="u4"
If you suspect the guard was triggered from C++, add TORCHDYNAMO_EXTENDED_DEBUG_CPP=1
For more debugging help, see https://docs.google.com/document/d/1HSuTTVvYH1pTew89Rtpeu84Ht3nQEFTYhAX3Ypa_xJs/edit?usp=sharing

For C++ stack trace, run with TORCHDYNAMO_EXTENDED_DEBUG_CPP=1

The following call raised this error:
  File "/home/ilyas/transformers/src/transformers/models/vits/modeling_vits.py", line 565, in forward
    hidden_states = self.conv_pre(first_half) * padding_mask

To fix the error, insert one of the following checks before this call:
  1. torch._check(inputs.shape[2] < 1)
  2. torch._check(inputs.shape[2] >= 1)

(These suggested fixes were derived by replacing `u4` with inputs.shape[2] or padding_mask.shape[2] or first_half.shape[2] or second_half.shape[2] in u4 < 1 and its negation.)

The error above occurred when calling torch.export.export. If you would like to view some more information about this error, and get a list of all other errors that may occur in your export call, you can replace your `export()` call with `draft_export()`.

[harshaljanjani]

That makes sense to me; thanks for the explanation of the revert and the context on the torch._check traceback as well :)