Reported by @fairydreaming here (thanks again!)

Custom extension introduced by HuggingFace in huggingface/transformers#30650

It looks like it's only useful during training (to help mask generated tokens away), so should be ok to treat as a no-op for inference purposes.

But... since it breaks jinja2, it may be time to switch the python expectations logic to the transformers library, which may simplify the tests. (Tracking in #30 )