Add a specific device_id when calling torch.init_process_group to avoid NCCL communication errors.

[gongweibao]

Such as

using GPU  to perform barrier as devices used by this process are currently unknown. 
This can potentially cause a hang if this rank to GPU mapping is incorrect. 
Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id

[ZeroRF]

It works! I have been confused by NCCL Error for a long time.