Fix XLA tensor storage device by using XlaDeviceToAtenDevice.

[ysiraichi]

Discussion: pytorch/pytorch#111506

This PR replaces backendDeviceToAtenDevice by XlaDeviceToAtenDevice when instantiating a new storage in the newly created XLA tensor.

Problem: there are 2 problems:

1. backendDeviceToAtenDevice doesn't take into consideration the device type
2. XLA uses XlaDeviceType enum as the device type, while eager uses c10::DeviceType

Solution: use XlaDeviceToAtenDevice function, since we know that the input is an XLA device anyway.

[ysiraichi]

@JackCaoG The failing CI doesn't seem to be related to this PR. Can we re-run it?

[JackCaoG]

Let me also give you write access so you can rerun CI too

[alanwaketan]

I wonder have you experienced any similar bugs on python level?

This is originally a hack see: https://github.com/pytorch/xla/blob/master/torch_xla/csrc/tensor.h#L334

We don't expect people to actually use storage_.

[ysiraichi]

As of now, we do hit that bug when running this example. That said, there are 2 things I'd like to clarify:

• We did decide to solve that bug, specifically, in some other way
• I still think we should be consistent with aten devices, here
  • e.g. XLATensorImpl also calls bridge::XlaDeviceToAtenDevice for constructing itself

[alanwaketan]

LGTM.

[alanwaketan]

As of now, we do hit that bug when running this example. That said, there are 2 things I'd like to clarify:

• We did decide to solve that bug, specifically, in some other way

• I still think we should be consistent with aten devices, here

  • e.g. XLATensorImpl also calls bridge::XlaDeviceToAtenDevice for constructing itself

Got it. Hopefully we can remove this workaround on fake storage soon.