Out-of-box XLA on Huggingface PyTorch models

[kevint324]

I'm not sure if this is the right place to ask since it involves multiple interest groups such as openxla community, pytorch/xla and huggingface.

Unlike tensorflow's effort to push XLA with HF, AFAIK there is little work on PyTorch/XLA and HF integration. And there is virtually zero model could work on XLA out-of-box, espacially with the latest LLMs with the recusive pattern usually trigger endless re-compilation.

Is out-of-box experience a goal or prioritized task of the openxla project?

Thanks