internalize _parse_xla_device

[zpcore]

Resolves #7652

Along with some nit fixes.

[will-cromar]

Would the deprecation wrapper here work? https://github.com/pytorch/xla/blob/master/torch_xla/experimental/deprecation.py

[zpcore]

Would the deprecation wrapper here work? https://github.com/pytorch/xla/blob/master/torch_xla/experimental/deprecation.py

Also did some updates to the deprecation decorator function. Previously it will print the warning message whenever we import torch_xla.

[will-cromar]

Would the deprecation wrapper here work? https://github.com/pytorch/xla/blob/master/torch_xla/experimental/deprecation.py

Also did some updates to the deprecation decorator function. Previously it will print the warning message whenever we import torch_xla.

This was actually not written as a decorator. It essentially aliases a function from the new location to the deprecated location. We don't need to keep the old implementation around in most cases. Check out some of the old usage that Jack removed in #7579

[zpcore]

Would the deprecation wrapper here work? https://github.com/pytorch/xla/blob/master/torch_xla/experimental/deprecation.py

Also did some updates to the deprecation decorator function. Previously it will print the warning message whenever we import torch_xla.

This was actually not written as a decorator. It essentially aliases a function from the new location to the deprecated location. We don't need to keep the old implementation around in most cases. Check out some of the old usage that Jack removed in #7579

Oh, I see. Just fixed wrapper it back.

[will-cromar]

Otherwise LGTM

[will-cromar]

Thanks! I think this should actually be torch_xla.core.xla_model looking at the implementation of register_deprecated. Please double-check that you get the expected message when you call torch_xla.core.xla_model.parse_xla-device(...)

[zpcore]

Updated to parse_xla_device = deprecated(torch_xla.core, iutils.parse_xla_device) instead.

[will-cromar]

The error message still says torch_xla.core.parse_xla_device instead of torch_xla.core.xla_model.parse_xla_device:

root@t1v-n-bf2f726f-w-0:/workspaces/ptxla# python
...
>>> import torch_xla.core.xla_model as xm
>>> xm.parse_xla_device('TPU:0')
WARNING:root:torch_xla.core.parse_xla_device is deprecated. Use torch_xla._internal.utils.parse_xla_device instead.
('TPU', 0)

You'll actually need to pass the xla_model module into register_deprecated. That was the only way I could find to print the qualified name of the aliased method consistently, but feel free to refactor it if you find a better way.

I think we need to add a unit test here to confirm that the right warning is getting printed, since you're going to be adding a lot of these aliases and this is an easy mistake to make. You can create a new unit test file for all of these public to internal aliases and we can remove it later.

[zpcore]

This should work:

import torch_xla._internal.utils as _utils
from . import xla_model as this_module
parse_xla_device = deprecated(this_module, _utils.parse_xla_device)

[will-cromar]

Thanks!