Fix tp error when torch distributed is already initialized

[SunMarc]

What does this PR do?

This PR fixes an issue when the torch distributed is already initialized. current_device is referenced before assignment.

Getting the following traceback in accelerate when testing tp training.

[rank1]:     tp_plan, device_map, device_mesh = initialize_tensor_parallelism(tp_plan, tp_size=None)
[rank1]:   File "/home/marc/transformers/src/transformers/integrations/tensor_parallel.py", line 79, in initialize_tensor_parallelism
[rank1]:     index = current_device.current_device() if device_type != "cpu" else None
[rank1]: UnboundLocalError: local variable 'current_device' referenced before assignment

[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.

[ArthurZucker]

thanks ! 🤗