remove use_xpu to fix ut issues, we don't need this since XPU is OOB …

[yao-matrix]

Sympton
When run transformers e2e training ut pytest -rA tests/trainer/test_trainer.py::TrainerIntegrationTest::test_end_to_end_example on XPU, we will get below error:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and xpu:0! (when checking argument for argument index in method wrapper_XPU__index_select)

Root Cause
When run this test case while there is no default accelerate config in “/root/.cache/huggingface/accelerate/default_config.yaml”（generated by accelerate config）, the use_xpu flag in launch.py (default False) blocked the using of XPU, so accelerate is launched both in single process and in cpu, but when data comes, it automatically detected xpu and put data into xpu, so model is in cpu and data is in xpu, lead to this error.

Fix

XPU is now native citizen of PyTorch, we need align the same behavior w/ CUDA, so remove these special handling.

[yao-matrix]

@sywangyi @faaany , this is a pretty fundamental change, pls review first

[yao-matrix]

@faaany , since it changes the default behavior, could you help run accelerate UTs for this? Thx

[faaany]

@faaany , since it changes the default behavior, could you help run accelerate UTs for this? Thx

sure

[faaany]

@yao-matrix no test error is found related to this PR.

[yao-matrix]

@SunMarc @muellerzr , pls help review, thx.

[muellerzr]

Overall on the surface: good. But let's be careful to deprecate please!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yao-matrix]

@muellerzr @SunMarc, updated per your comments. Pls help review whether any other changes needed.

[SunMarc]

Thanks !