Remove all hf-internal-testing checkpoints that can be removed

[sgugger]

What does this PR do?

This PR continues the work on docstrings and removes all checkpoints from the hf-internal-testing org where they can be removed.

[HuggingFaceDocBuilderDev]

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

[ydshieh]

Thanks for working on this. I believe there are 2 places to correct, and one for you to decide (reformer)

[ydshieh]

I believe you intended to change the file modeling_longformer.py instead of this TF file. As the TF_xxx_SAMPLE still have something like (at this moment)

from transformers import {processor_class}, {model_class}

[sgugger]

Ah yes, thanks!

[ydshieh]

I think this could use _CHECKPOINT_FOR_DOC but remove the parts that check the expected output values, and we don't need this warning?

At least, this is what you have done for some models that used tiny-random-xxx checkpoints.

[ydshieh]

(Hmm, it raises a question: when we use real checkpoints which don't contain head weights, so far we don't include the output checks, but we don't do any warning. The warning introduced in your last PR is only used for 2 models so far where no small enough real checkpoints exist).

Would like to hear from you regarding this - as I remember one major reason is to avoid user confusion regarding the results they get.

[sgugger]

Here the checkpoint fails because it has is_decoder=True and we need False to use the MLM head.

[ydshieh]

OK, thanks for explaining :-)

[ydshieh]

Not able to comment at the exact line, but for UniSpeechSatForPreTraining, I think you missed to change Wav2Vec2FeatureExtractor (you changed this for another model file UniSpeechForPreTraining)

[ydshieh]

LGTM + thank you again!