if args.local_rank not in [-1, 0]:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

        ... (loads the model and the vocabulary)

    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

 理解：

  四个进程（0，1,  2 ，3）多卡训练模型时都是同步并行的，但是在读取数据，数据预处理等操作是不需要并行做的。一般只需要主进程（local_rank = 0）进行这些操作。

在执行到第一个if语句，其他进程（local_rank != 0）会被阻塞。主进程执行后面操作。直到执行第二个if语句时，主进程也被阻塞。当所有进程都被阻塞时，torch.distributed.barrier()会释放所有进程。

参考：GitHub - jia-zhuang/pytorch-multi-gpu-training: 整理 pytorch 单机多 GPU 训练方法与原理整理 pytorch 单机多 GPU 训练方法与原理. Contribute to jia-zhuang/pytorch-multi-gpu-training development by creating an account on GitHub.https://github.com/jia-zhuang/pytorch-multi-gpu-training

 pytorch - How does torch.distributed.barrier() work - Stack Overflowhttps://stackoverflow.com/questions/59760328/how-does-torch-distributed-barrier-work

Distributed communication package - torch.distributed — PyTorch 1.11.0 documentation