A Dockerfile for building CUDA enabled apps. It was a pain getting Pytorch to run from a container for an app and use GPUs at inference time in the configuration I needed. So, once I figured it out I figured there might be some other people looking for something similar. Essentially, create your folder you want to turn into a container, make your requirements.txt file, make sure the FROM at the start of the Dockerfile has the correct Ubuntu and CUDA version for your application. Build the app, then run it with docker run --runtime=nvidia imagename and everything should work.