How can I use NCCL2 backend for distributed training with v0.4.0?

I want to use nccl2 as distributed backend.
I just downloaded nccl2 from nccl nvidia download page.

  • nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb
  • nccl-repo-ubuntu1604-2.2.13-ga-cuda8.0_1-1_amd64.deb

Please provide some instructions for using nccl backend for DistributedDataParallel.

1 Like