The pytorch program exits when execute after DistributedDataParallel

Hi all,

I am currently using apex for multi-gpu training with pytorch 1.01. I compile the apex successfully, however, the training program exit when it executes at this line without any information:

from apex.parallel import DistributedDataParallel, SyncBatchNorm

print('hello 93')
if engine.distributed:
    model = DistributedDataParallel(model) #the program exit at this line
print('hello 96')

I'm sure the model (network) is fine. I'd like to ask what's the possible problem and any suggestions?

Environment:
pytorch 1.01
apex 0.1
cuda 9 (I have also tried cuda 10)

Thanks.

Start at apex.parallel.DistributedDataParallel and review how it is invoked in the reported multi-GPU setup, including the surrounding distributed initialization. Use the stated PyTorch 1.01, Apex 0.1, and CUDA environments when investigating; done means identifying a reproducible cause for the silent exit or documenting the missing information needed to diagnose it.