Next batch of models with removed return_dict #37396
Conversation
|
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. |
| output_attentions: Optional[bool] = None, | ||
| output_hidden_states: Optional[bool] = None, | ||
| return_dict: Optional[bool] = None, |
There was a problem hiding this comment.
Not super related to the PR, but I removed redundant kwargs for get_text_features/get_image_features, which actually has no effect because we return a tensor (not Output) anyway, might be worth adding 🚨🚨🚨 for the PR
| output_attentions=False, | ||
| output_hidden_states=False, |
There was a problem hiding this comment.
passing False explicitly here instead (see above comment)
| @@ -463,7 +463,6 @@ class BertModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin | |||
| if is_torch_available() | |||
| else {} | |||
| ) | |||
| fx_compatible = True | |||
There was a problem hiding this comment.
FX compatibility is broken for BERT, ELECTRA, MobileBERT, and RoBERTa. I'm not sure how critical this is. As far as I understand, we dropped FX compatibility for text models previously by adding **kwargs.
It's a bit strange that FX compatibility is not broken for CLIP, even though it also has the @can_return_tuple decorator.
What does this PR do?
Continue removing the
return_dictattribute from the modeling code.