distributed training compatible issue in ignite 0.4.2

[Nic-Ma]

❓ Questions/Help/Support

Hi @vfdev-5 ,

I am trying to upgrade ignite to v0.4.2 in MONAI, got error when I ran this test program of MONAI:
https://github.com/Project-MONAI/MONAI/blob/master/tests/test_handler_rocauc_dist.py
I used 2 GPU in 1 node, and it passed in ignite v0.3.0 before.
Here is the error log:

root@apt-sh-ai:/workspace/data/medical/MONAI# python -m torch.distributed.launch --nproc_per_node=2 --nnodes=1 --node_rank=0 --master_addr="10.23.137.29" --master_port=1234 tests/test_handler_rocauc_dist.py
*****************************************
Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
*****************************************
Traceback (most recent call last):
Traceback (most recent call last):
  File "tests/test_handler_rocauc_dist.py", line 48, in <module>
  File "tests/test_handler_rocauc_dist.py", line 48, in <module>
    main()
    main()
  File "tests/test_handler_rocauc_dist.py", line 23, in main
  File "tests/test_handler_rocauc_dist.py", line 23, in main
    auc_metric = ROCAUC(to_onehot_y=True, softmax=True)
    auc_metric = ROCAUC(to_onehot_y=True, softmax=True)
  File "/workspace/data/medical/MONAI/monai/handlers/roc_auc.py", line 66, in __init__
  File "/workspace/data/medical/MONAI/monai/handlers/roc_auc.py", line 66, in __init__
    super().__init__(output_transform, device=device)
  File "/opt/conda/lib/python3.6/site-packages/ignite/metrics/metric.py", line 200, in __init__
    super().__init__(output_transform, device=device)
  File "/opt/conda/lib/python3.6/site-packages/ignite/metrics/metric.py", line 200, in __init__
    if idist.get_world_size() > 1:
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/utils.py", line 133, in get_world_size
    if idist.get_world_size() > 1:
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/utils.py", line 133, in get_world_size
    sync(temporary=True)
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/utils.py", line 64, in sync
    sync(temporary=True)
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/utils.py", line 64, in sync
    model = comp_model_cls.create_from_context()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 48, in create_from_context
    model = comp_model_cls.create_from_context()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 48, in create_from_context
    return _NativeDistModel()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 64, in __init__
    return _NativeDistModel()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 64, in __init__
    self._init_from_context()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 97, in _init_from_context
    self._init_from_context()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 97, in _init_from_context
    self._setup_attrs()
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/base.py", line 26, in _setup_attrs
    self._setup_attrs()
    self._nproc_per_node = self._compute_nproc_per_node() if self.get_world_size() > 1 else 1
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/base.py", line 26, in _setup_attrs
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 101, in _compute_nproc_per_node
    self._nproc_per_node = self._compute_nproc_per_node() if self.get_world_size() > 1 else 1
  File "/opt/conda/lib/python3.6/site-packages/ignite/distributed/comp_models/native.py", line 101, in _compute_nproc_per_node
    dist.all_reduce(tensor, op=dist.ReduceOp.MAX)
  File "/opt/conda/lib/python3.6/site-packages/torch/distributed/distributed_c10d.py", line 938, in all_reduce
    dist.all_reduce(tensor, op=dist.ReduceOp.MAX)
  File "/opt/conda/lib/python3.6/site-packages/torch/distributed/distributed_c10d.py", line 938, in all_reduce
    work = _default_pg.allreduce([tensor], opts)
RuntimeError: NCCL error in: ../torch/lib/c10d/ProcessGroupNCCL.cpp:558, invalid usage, NCCL version 2.7.8
    work = _default_pg.allreduce([tensor], opts)
RuntimeError: NCCL error in: ../torch/lib/c10d/ProcessGroupNCCL.cpp:558, invalid usage, NCCL version 2.7.8
Traceback (most recent call last):
  File "/opt/conda/lib/python3.6/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/opt/conda/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/opt/conda/lib/python3.6/site-packages/torch/distributed/launch.py", line 261, in <module>
    main()
  File "/opt/conda/lib/python3.6/site-packages/torch/distributed/launch.py", line 257, in main
    cmd=cmd)
subprocess.CalledProcessError: Command '['/opt/conda/bin/python', '-u', 'tests/test_handler_rocauc_dist.py', '--local_rank=1']' returned non-zero exit status 1.

Something wrong with my NCCL version & ignite v0.4.2?

Thanks.

[vfdev-5]

@Nic-Ma thanks for the report ! Let me check it from my side to understand where is the problem.

[vfdev-5]

@Nic-Ma could you please provide more detail on your system: pytorch version and how it is built with NCCL ?

I can not reproduce your issue with NCCL. In my case, I use python 1.6.0 with prebuilt NCCL

torch.cuda.nccl.version()
> 2408

However, there can be another issue with the test which might be missing

torch.cuda.set_device("cuda:{}".format(local_rank))

This can lead to hangs on collective ops like all_gather. See https://pytorch.org/docs/stable/distributed.html#launch-utility (Important Notices).

This problem is emphasized with 0.4.2 as it implicitly relies on on the fact that user sets cuda device per local rank. However, I'd like to improve this part of our code.

[Nic-Ma]

Hi @vfdev-5 ,

Thanks for your quick help, after adding torch.cuda.set_device, issue solved.
I think maybe you can add some explicit warning if missing this setting as 0.4.2 implicitly relies on this setting.

Thanks.

[vfdev-5]

@Nic-Ma yes, this should be definitely addressed in a clear way ! Thanks for the suggestion !

I'm still wondering which pytorch distribution you are using with NCCL 2.7.8 ?
I think, it would be helpful for us to use it inside a CI.

[Nic-Ma]

Hi @vfdev-5 , we used this docker: https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
I think it's PyTorch 1.7a.
Thanks.

[vfdev-5]

Let's close this issue in favor of #1308