Question about loading the model that was trained using 4GPU with distributed dataparallel to only 1 GPU job

Hi :smiley: , I’m having trouble loading the distributed dataparallel model to just 1 GPU. And I want to know how to load the model (trained by 4 GPU with distributed dataparallel) to another job using only 1 GPU.

I have trained a model using 4 GPU and distributed dataparallel, and I saved it as the tutorial:
However, I don’t know how to load it using just 1 GPU for some simple job like validation test.

if rank == 0:, CHECKPOINT_PATH)

I’m now using this method:

    # initialize
    local_rank = torch.distributed.get_rank()
    device = torch.device("cuda", local_rank)
    # only gpu with rank0 can remain running

    model = resnet50()
    model = torch.nn.parallel.DistributedDataParallel(model, 
    if local_rank == 0:
        acc, acc_std, th = lfw_test(model, cfg.TEST.lfw_root, cfg.TEST.lfw_test_list)

and the command code:

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4

This method works, and through nvidia-smi I saw that only GPU0 is working, but when I run another test process using GPU device 1 (when the previous one is still running):

CUDA_VISIBLE_DEVICES=1,2,3 python -m torch.distributed.launch --nproc_per_node=3

The previous process throw a runtime error:

RuntimeError: NCCL error in: /opt/conda/conda-bld/pytorch_1549633347309/work/torch/lib/c10d/ProcessGroupNCCL.cpp:260, unhandled system error

There are 2 reasons for me to load the model using 1GPU:

  1. Some jobs have file-writing part and distributed parallel may cause wrong order.
  2. Running 4 tiny experiment with 1 GPU per process is more efficient for me to test my idea and finding bugs.

So is there a way that I can load the model like the common ways :torch.load_state_dict(torch.load()).to(torch.device("cuda:0))? :smile:

Oh… the first method is not work, neither. I find that using:

if local_rank == 0:
    output = model(input)

the model(input) will never output. And the code is just blocked there.

I’m not sure to understand the use case.
It seems you would like to load the state_dict to a single GPU machine, but in your code you are wrapping the model again in DDP.

Would creating the model, loading the state_dict, and pushing the model to the single GPU not work?

1 Like

It works! Thank you :smile:

I used:

model = resnet50()

and I got a Runtime error:

RuntimeError: Error(s) in loading state_dict for Resnet:
    Missing key(s) in state_dict: "conv1.weight", "bn1.weight" ... ...
    Unexpected key(s) in state_dict: "module.conv1.weight", "module.bn1.weight" ... ...

Seems the Distributed DataParallel save the model in module. Then I find a solution there:

    state_dict = torch.load(cfg.MODEL.pretrained_model_path)
    new_state_dict = OrderedDict()
    for k, v in state_dict.items():
        name = k[7:]
        new_state_dict[name] = v