Got stuck when training with multiple GPU using dist_train.sh

[xiazhongyv]

All child threads getting stuck when training with multiple GPU using dist_train.sh
With CUDA == 11.3, Pytorch == 1.10
After diagnosis, I found it was stuck at https://github.com/open-mmlab/OpenPCDet/blob/master/pcdet/utils/common_utils.py#L166-L171

I modified the code from

dist.init_process_group(
        backend=backend,
        init_method='tcp://127.0.0.1:%d' % tcp_port,
        rank=local_rank,
        world_size=num_gpus
)

to

dist.init_process_group(
        backend=backend
)

and it worked.

I'm curious why this is so, and if someone else is having the same problem, you can try to do the same.

[dk-liang]

Thanks. I have the same problem, and I solved it using your method.

[Eaphan]

@sshaoshuai After you fix bug in this way, the tcp_port is not used actually.
Can you fix it in a more decent way?

[sshaoshuai]

Thank you for the bug report. It has been fixed in #784.

Can you help to double check whether it works now?

[Eaphan]

Thank you for the bug report. It has been fixed in #784.

Can you help to double check whether it works now?

@sshaoshuai Thanks for your work. It's ok now.

[aotiansysu]

For single-machine multi-GPU training, I also modified the local_rank to rank in torch.cuda.set_device() to be able to train properly. Otherwise it throws this error: Duplicate GPU detected : rank 0 and rank 1 both on CUDA device a000.
Modified:

def init_dist_pytorch(tcp_port, local_rank, backend='nccl'):
    if mp.get_start_method(allow_none=True) is None:
        mp.set_start_method('spawn')
    num_gpus = torch.cuda.device_count()

    dist.init_process_group(
        backend=backend,
    )

    rank = dist.get_rank()
    torch.cuda.set_device(rank % num_gpus)
    return num_gpus, rank

[jiaminglei-lei]

@sshaoshuai
torch=1.9.0 cuda=11.1.
Got stuck at dist.init_process_group and the code is latest....
In other distribued training project having the same code for init_process_group, it ran successfully. ......

[jiaminglei-lei]

@sshaoshuai torch=1.9.0 cuda=11.1. Got stuck at dist.init_process_group and the code is latest.... In other distribued training project having the same code for init_process_group, it ran successfully. ......

after I uncomment the lines mentioned in #784 (comment), it works.

[sshaoshuai]

I have submitted a new PR to solve this issue in #815.

Please pull the latest master branch if you still get block when training with dist_train.sh.