Support ONNX export for causal LM sequence classifiers

[dwyatte]

What does this PR do?

Partial fix for huggingface/optimum#1527 in optimum when exporting causal LMs with sequence classification support to ONNX

ONNX's argmax operator does not support int64, but that should not be needed here since these are just boolean tensors

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 write any new necessary tests?

Who can review?

@ArthurZucker and @younesbelkada (CC @fxmarty)

[fxmarty]

Thank you for the fix @dwyatte, LGTM! Feel free to open a PR in optimum as well to re-enable the export of those models for text-classification.

[ArthurZucker]

Nice followup to #24979. Not really sure why I did not use int at the time, but I ran the slow tests and this seems to be alright! Thanks 🤗

[ArthurZucker]

The failing test might just need a rebase to main otherwise I'll skip it on main and work on a fix

[amyeroberts]

Thanks for fixing!

[dwyatte]

The failing test might just need a rebase to main otherwise I'll skip it on main and work on a fix

@ArthurZucker I rebased in 8076250 but looks like something is still up with CI. Perhaps different tests get selected based on files changed or between PRs/main

• torch_tests from this PR (14627 tests selected): https://app.circleci.com/pipelines/github/huggingface/transformers/78297/workflows/e34f774a-5910-488e-942d-7121d1007bf1/jobs/998415
• torch_tests on main 78f6ed6 (7619 tests selected): https://app.circleci.com/pipelines/github/huggingface/transformers/78269/workflows/3e9cb27d-42b0-4b00-bf5f-432eec81c4c0/jobs/997963

[amyeroberts]

@dwyatte Yes, the test fetcher selects a subset of the tests to run based on the files that are touched. In this case, the failing tests (I believe) are unreleated to your PR. The tests involving safetensors have had a patch pushed on main. Could you rebase on main to include these in the test runners?

[dwyatte]

@amyeroberts I think there are some other problems on main still. Here's what's failing in the tests for this PR after the rebases in 8076250 and d8ab2c9

FAILED tests/models/switch_transformers/test_modeling_switch_transformers.py::SwitchTransformersModelTest::test_assisted_decoding_sample - RuntimeError: The size of tensor a (3) must match the size of tensor b (4) at non-singleton dimension 3
FAILED tests/models/t5/test_modeling_t5.py::T5ModelTest::test_assisted_decoding_sample - RuntimeError: The size of tensor a (3) must match the size of tensor b (4) at non-singleton dimension 3
FAILED tests/models/speech_to_text/test_modeling_speech_to_text.py::Speech2TextModelTest::test_tf_from_pt_safetensors - AssertionError: False is not true

[VsonicV]

@dwyatte Exactly same unrelated CI failures in #27351 . In addition to the failures related to safetensors (you already mentioned above), we also need to resolve the other CI failures caused by test_assisted_decoding_sample in tests_torch.

[ArthurZucker]

Sorry both for the delays, I'll skip these 3 tests as well. cc @gante I'll look into the test_assisted_decoding_sample.

[VsonicV]

Hi, @ArthurZucker , regarding the failures caused by test_assisted_decoding_sample, there have already been some discussions in #26892 , these failures did not only happen for switch_transformers and t5, but also happened for blenderbot , pegasus and umt5 in my previous CI checks. It seems to be a more general issue that may happen to more than these listed models. Thanks for the help in fixing them!

[ArthurZucker]

just merged #27508 which should skip it for all models

[dwyatte]

Thanks @ArthurZucker that took care of the remaining failures. This is ready to merge