Fix TF CTC tests

[gante]

What does this PR do?

(see title)

[gante]

ninja copy paste hehe

[gante]

This one, in TFHubertModelTest, was the issue. Since it did not have the lower batch size treatment as in the other TF CTC tests, I'm assuming this is the fix.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ydshieh]

OK for me. But would like @amyeroberts to take a look (even afterward) to check if any merged commit cause this issue. These tests run with tiny models, usually don't have such memory issue. And we didn't have this failure previously (right?).

[gante]

@ydshieh before @amyeroberts work in #21502 these tests had no overwrite, correct. However, see the note in Amy's PR -- the labels were not correctly handled, which meant that test_dataset_conversion was probably terminating early here, which would explain the ausence of crashes prior to the PR :)

[sgugger]

Thanks for the fix!

[amyeroberts]

Thanks for the fix @gante !

@ydshieh Yes, Joao's correct. There's three different cases where the state of the tests changed:

• For test_loss_computation, the previous test was being skipped. The given reason was incorrect shapes - however, at the time of adding Replace input_values_processing with unpack_inputs #21502 the returned loss was actually None.
• Some tests were being skipped by overloading with an empty test rather than unittest.skip and so previously showed as passing.
• The branch calculating loss wasn't being touched when fitting the model - input_values passed in as a dictionary - as labels was taken from keyword arguments, rather than outputs['labels']. This meant some tests e.g. test_keras_fit in TFModelTesterMixin previously passed as the memory intensive operation was skipped.