aavbsouza

Hello @xiaogengyaokeyan .This is a expected behavior since Horovod uses multiple processes. You should use some other tool for instance this solution from PFN (https://github.com/pfnet/pytorch-pfn-extras/blob/master/pytorch_pfn_extras/dataset/shared_dataset.py) or libraries like Ray Dataset.

Hello @samcmill. I have changed the PR to also work with singularity and bash containers. For the Singularity case I have used your suggestion for the bash syntax to define...

Hello @samcmill any toughs on the current state of the PR ?

Hello @samcmill sorry for the delay. I will update the pull request. thanks

I believe this is the expected behavior, due to the use of variables like `NVIDIA_VISIBLE_DEVICES` and also is common in frameworks like pytorch as in `torch.cuda.current_device()` that also would return...