Nvidia Docker images

Nvidia has a repo of all docker images it creates, one of them: Torch | NVIDIA NGC


“Das finde ich zielführender als…” - heard at work

Docker - automatically assign a free port

docker run --name frontend -p 0:80 frontend:latest1

Port 0 gets passed to the kernel that assigns any free port.

To see which one, docker port somecontainer.

Docker run container on specific GPU

docker run --gpus device=3 -e NVIDIA_VISIBLE_DEVICES=0 -e CUDA_VISIBLE_DEVICES=0 myservice

Where the device=3 is the GPU id on the host that we want to use.