For xla tensors, use an alternative way to get a unique id

[qihqi]

What does this PR do?

XLA tensors don't have storage and attempting to get it's storage will get

RuntimeError: Attempted to access the data pointer on an
invalid python storage.

Repro:

from transformers import pytorch_utils
import torch_xla.core.xla_model as xm
device = xm.xla_device()
a = torch.ones((10,10)).to(device)
pytorch_utils.id_tensor_storage(a)  # raises RuntimeError

With this patch it would print out
(device(type='xla', index=0), 1, 400).

[qihqi]

Hi @muellerzr moved huggingface/safetensors#349 to here

[qihqi]

LMK if I should add a test (and in which file) I am not sure if the CI machines has torch_xla installed so import torch_xla can succeed...

[muellerzr]

Thanks! Some general feedback on how we check torch_tpu, cc @amyeroberts

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[amyeroberts]

Thanks for this PR!

Changes look OK to me. However, it feels contradictory to add a work around for tensors that don't have storage within a functional called id_tensor_storage: this function shouldn't be called on those tensors at all.

@muellerzr @qipeng Could you provide some more context for this change? i.e. is this enabling functionality elsewhere?

[muellerzr]

@amyeroberts this stems from huggingface/safetensors#349, where we specified that it should be done in transformers and accelerate separately.

[muellerzr]

As long as @amyeroberts is happy with the explanation, lg2m

[amyeroberts]

Thanks for adding this fix!