How to take multi-GPU support to the OpenNMT-py (pytorch)?

Question

I used python-2.7 version to run the PyTorch with GPU support. I used this command to train the dataset using multi-GPU. Can someone please tell me how can I fix this error with PyTorch in OpenNMT-py or is there a way to take pytorch support for multi-GPU using python 2.7?

Here is the command that I tried.

CUDA_VISIBLE_DEVICES=1,2 python train.py -data data/demo -save_model demo-model -world_size 2 -gpu_ranks 0 1

This is the error:

Traceback (most recent call last): File "train.py", line 200, in main(opt) File "train.py", line 60, in main mp = torch.multiprocessing.get_context('spawn') AttributeError: 'module' object has no attribute 'get_context'

Answer 1

Maybe you can check whether your torch and python versions fit the openmt requiremen. I remember their torch is 1.0 or 1.2 (1.0 is better). You have to lower your latest of version of torch. Hope that would work