How to torch.cuda.set_device with torch.distributed.launch

Hi all,
I’m trying to use torch.distributed.launch with NCCL backend on two nodes each of them has single GPU. When I see here, it guides me to set torch.cuda.set_device(local_rank), however, each node has only device 0 available. So I’m confused torch.cuda.set_device(0) for both process is correct or not.
Either of them I met an error like this:

Traceback (most recent call last):
  File "", line 26, in <module>
  File "/u3/jbaik/pytorch-asr/asr/models/deepspeech_ctc/", line 56, in batch_train
    trainer = NonSplitTrainer(model, **vars(args))
  File "/u3/jbaik/pytorch-asr/asr/models/", line 93, in __init__
    self.model = nn.parallel.DistributedDataParallel(model, device_ids=[local_rank], output_device=local_rank)
  File "/home/jbaik/.pyenv/versions/3.7.0/lib/python3.7/site-packages/torch/nn/parallel/", line 134, in __init__
  File "/home/jbaik/.pyenv/versions/3.7.0/lib/python3.7/site-packages/torch/nn/parallel/", line 251, in _dist_broadcast_coalesced
    dist.broadcast(flat_tensors, 0)
  File "/home/jbaik/.pyenv/versions/3.7.0/lib/python3.7/site-packages/torch/distributed/", line 279, in broadcast
    return torch._C._dist_broadcast(tensor, src, group)
RuntimeError: NCCL error in: /u3/setup/pytorch/pytorch/torch/lib/THD/base/data_channels/DataChannelNccl.cpp:322, unhandled system error
1 Like

If each node has multiple NIC, does NCCL finds the proper connection between the nodes? How about the other backends?

In my case, the same error was found when I used docker.

With ‘–network=host’ parameter, the problem was resolved.

Hi, is --network=host parameter is in nvidia-docker run command in all nodes?