Fix ProphetNet forward to handle tuple encoder_outputs

[Abdennacer-Badaoui]

This PR fixes a bug in ProphetNet where passing encoder_outputs as a tuple to forward() would raise:

AttributeError: 'tuple' object has no attribute 'last_hidden_state'

This happens because the forward() method assumes encoder_outputs is a BaseModelOutput when provided manually, while the docstring says it can be a tuple.

Changes:

Added a check in forward():

Convert tuple encoder_outputs to BaseModelOutput format before passing to ProphetNetSeq2SeqModelOutput.

if isinstance(encoder_outputs, tuple):
    encoder_outputs = BaseModelOutput(
        last_hidden_state=encoder_outputs[0],
        hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
        attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
    )

[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.

[Rocketknight1]

cc @ArthurZucker for text models

[ArthurZucker]

I think we'd rather update the docstring as it should no longer be a tuple!

[Abdennacer-Badaoui]

Alright! In this case, the ProphetNetEncoder should be modified to always return a BaseModelOutput. This means removing the return_dict argument, as it will not be used there.

[ArthurZucker]

btw if you have time, can you add the check_model_input decorator, updating to the latest standards? 🤗

[ArthurZucker]

sorry, what I mean by that is add this, remove the args, add the TransformersKwargs, remove the

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

etc! Have a look at modeling llama to see how we handle ti

[Abdennacer-Badaoui]

My bad, it’s clearer now. I’ll try to fix it. Thanks!

[github-actions]

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

run-slow: prophetnet