Prolly why I don’t like Huggingface very much - its just too fragmented for useful customization (though to its credit, it does warn about that con in the readme). I would be getting a host of new errors now with new scripts 
- Why do we need to use the Flax version for running BigBird on TPU
- Why did Google opt to release BigBird on HF, rather than a standalone Pytorch/Jax repo?