training gets frozen while using multiple-GPUs

[NathanYanJing]

Super cool and amazing work!

I am writing to ask for your assistance with an issue I am encountering while training a model using A6000 GPUs. I am using the following command to run my code:

torchrun --nnodes=1 --nproc_per_node=4  train.py --model DiT-XL/2 --data-path training_data --global-batch-size 76 --num-workers 1

The problem I am experiencing is that the training appears to be frozen after creating the experiment directory for a long period of time. On occasion, it also throws the following error:

Traceback (most recent call last):
  File "/DiT/DiT/train.py", line 269, in <module>
[E ProcessGroupNCCL.cpp:456] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
[E ProcessGroupNCCL.cpp:461] To avoid data inconsistency, we are taking the entire process down.
    main(args)
  File "/DiT/DiT/train.py", line 149, in main
    model = DDP(model.to(device), device_ids=[rank])
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/nn/parallel/distributed.py", line 655, in __init__
    _verify_param_shape_across_processes(self.process_group, parameters)
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/utils.py", line 112, in _verify_param_shape_across_processes
    return dist._verify_params_across_processes(process_group, tensors, logger)
RuntimeError: DDP expects same model across all ranks, but Rank 0 has 131 params, while rank 1 has inconsistent 0 params.
[E ProcessGroupNCCL.cpp:456] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
[E ProcessGroupNCCL.cpp:461] To avoid data inconsistency, we are taking the entire process down.
Traceback (most recent call last):
  File "/DiT/DiT/train.py", line 269, in <module>
    main(args)
  File "/DiT/DiT/train.py", line 149, in main
    model = DDP(model.to(device), device_ids=[rank])
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/nn/parallel/distributed.py", line 655, in __init__
    _verify_param_shape_across_processes(self.process_group, parameters)
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/utils.py", line 112, in _verify_param_shape_across_processes
    return dist._verify_params_across_processes(process_group, tensors, logger)
RuntimeError: [2]: params[0] in this process with sizes [1152, 4, 1, 1] appears not to match sizes of the same param in process 0.
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 225631 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 225633 closing signal SIGTERM
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 225634 closing signal SIGTERM
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: -6) local_rank: 1 (pid: 225632) of binary: /miniconda3/envs/DiT/bin/python
Traceback (most recent call last):
  File "/miniconda3/envs/DiT/bin/torchrun", line 33, in <module>
    sys.exit(load_entry_point('torch==1.13.1', 'console_scripts', 'torchrun')())
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
    return f(*args, **kwargs)
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/run.py", line 762, in main
    run(args)
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/run.py", line 753, in run
    elastic_launch(
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/miniconda3/envs/DiT/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 246, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

I have not experienced this problem when training with 1, 2, or 3 nodes.

I apologize for my lack of experience in this area, but could you please provide any insights or guidance to help me resolve this issue? Thank you for your assistance.

[NathanYanJing]

Problems solved for now! In case people might encounter a similar issue, If you use single node multiple GPU, replace the DDP with the following, there is a hacky way, from torch.nn import DataParallel as DDP Or you can try the following torch.multiprocessing.set_start_method('spawn',force=True) but you might need to rewrite the lambda function to avoid the pickle issue.

[wpeebles]

Hi @NathanYanJing. Your torchrun command runs fine for me without any modifications to the code (also using a single-node, multi-GPU training setup). I haven't run across the error you're getting before. Depending on how you're launching the script, you might want to be a little careful with the DDP --> DataParallel change since that could change the behavior of parts of train.py that rely on distributed ops (in general I'm not sure if DataParallel plays nice with torch.distributed)

[NathanYanJing]

Hi @wpeebles , thanks for your reply! Yea, I agree with using torch.distributed is always a better choice.

Yes, it seems the problem comes back again somehow now -- it hangs at Dataloader part. I am guessing that this is probably NCCL and Nvidia-version issue. Would you mind sharing your NCCL and Cuda versions?