Effectively allow encoder_outputs input to be a tuple in pix2struct

[fxmarty]

Be consistent with the type hint

transformers/src/transformers/models/pix2struct/modeling_pix2struct.py

Line 1656 in fabe17a

encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,

(which should rather be Optional[Union[Tuple[Tuple[torch.FloatTensor]], OrderedDict]] IMO) and follow what is done with other architectures.

Otherwise, currently, with return_dict=True, an error is raised as we later try to access encoder_outputs.last_hidden_state:

transformers/src/transformers/models/pix2struct/modeling_pix2struct.py

Lines 1768 to 1781 in fabe17a

	if not return_dict:
	return decoder_outputs + encoder_outputs
	return Seq2SeqLMOutput(
	loss=decoder_outputs.loss,
	logits=decoder_outputs.logits,
	past_key_values=decoder_outputs.past_key_values,
	decoder_hidden_states=decoder_outputs.hidden_states,
	decoder_attentions=decoder_outputs.attentions,
	cross_attentions=decoder_outputs.cross_attentions,
	encoder_last_hidden_state=encoder_outputs.last_hidden_state,
	encoder_hidden_states=encoder_outputs.hidden_states,
	encoder_attentions=encoder_outputs.attentions,
	)

This is blocking for the ONNX export support.

[HuggingFaceDocBuilderDev]

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

[younesbelkada]

Thank you for fixing!

[sgugger]

Thanks!