RuntimeError: Default process group has not been initialized, please make sure to call init_process_group.

    if args.local_rank == -1:
        device = torch.device("cuda")
    else:
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        deepspeed.init_distributed()
    args.global_rank = torch.distributed.get_rank()

     此处代码是不是有问题？单独跑train.py 报错的，不用命令跑，不能单独调试代码，只能用distribute模式
