# Transformer Lens + NNsight

$ TypeError: HookedTransformer.forward() got an unexpected \
keyword argument 'input_ids'

- `.forward()` method in Transformer Lens uses `input` kwarg. LanguageModel in NNsight uses `input_ids` to prepare inputs.

### Fix

- Check function signature in NNsightModel or make a pull request to change TransformerLens naming conventions.

<br/><br/>

$ TypeError: HookedTransformer.forward() got an unexpected \
keyword argument 'labels'

- Some auto transformer models use `labels` to compute loss at the end. TransformerLens does not support these models.

### Fix

def _example_input(self) -> Dict[str, torch.Tensor]:
        meta_signature = inspect.signature(self.meta_model.forward) 

        if "labels" in meta_signature.parameters.keys():
            return BatchEncoding(
                {"input": torch.tensor([[0]]), "labels": torch.tensor([[0]])}
            )
        else:
            return BatchEncoding({"input": torch.tensor([[0]])})

- Use `inspect.signature()` to check if model forward method takes labels.

# Summary

- Might be better just to include a wrapper with NNsight.