XLA error on CPU

[ckchow]

https://github.com/PyTorchLightning/pytorch-lightning/blob/6e124e7207f6459cb43f540cfb5a1c6cc9b00f7a/pytorch_lightning/utilities/xla_device.py#L75

A somewhat unusual case, but this line causes a TypeError if you are using an XLA-enabled version of pytorch without any TPUs attached. (Google deprecated all of its non-XLA pytorch deep learning images so maybe this will become less unusual over time.)

This is because pytorch/xla returns a None and not an empty list if it can't find any devices of the desired type. Could we guard this with an is None or?

[SeanNaren]

This sounds reasonable, would you be able to make a PR for this? cc @kaushikb11 any ideas how we'd test this?

[ckchow]

I got it to pop by using the docker image gcr.io/deeplearning-platform-release/pytorch-xla.1-9, installing pytorch-lightning>=1.4.2, and I think just importing Trainer.