Remove unused token_type_ids in MPNet #9564
Conversation
There was a problem hiding this comment.
cc @LysandreJik But I don't think this is the way to go. We should just silently remove the token_type_ids from the tokenizer and the model args at the same time. I don't think this warrants a deprecation cycle since it's not used.
|
I'm totally fine to definitely suppress this argument in once if this is prefered (I would prefer as well) |
There was a problem hiding this comment.
I don't think we need to put warnings however, as @sgugger said I would simply remove these from both the tokenizer and the forward call. Their presence is a bug, and not a feature we're modifying.
| @@ -582,6 +578,7 @@ def call( | |||
| training=training, | |||
| kwargs_call=kwargs, | |||
| ) | |||
| print(inputs) | |||
jplu
changed the title
|
Not an expert neither. Maybe @LysandreJik knows better. |
I think it's still required as it puts e.g. the [sep] token correctly between two sentences. I don't think that |
Co-authored-by: Lysandre Debut <[email protected]>
What does this PR do?
This PR adds a warning when the argument
token_type_idsis given a show a message saying that this argument is never used. I just supressed the internal appearance of this argument without modifying the method signatures in order to do not integrates a breaking change.Should I update the Tokenizer to make it returns only
attention_mask?