Usage

direct call

session 1

$ CUDA_VISIBLE_DEVICES=1,3 python remote_train.py --rank 0 --world_size 2

session 2

$ CUDA_VISIBLE_DEVICES=1,3 python remote_train.py --rank 1 --world_size 2

---

run with Docker

Note: nvidia-docker has to be installed

$ docker build --file Dockerfile --tag mini .
$ CUDA_VISIBLE_DEVICES=1,3 docker-compose --file docker-compose-m2.yml up