Reduce image size with multi-stage builds

[matthewfeickert]

It may be possible to reduce the image sizes with multi-stage builds, though this might take lots of effort.

Some things that might be worth reading after a Google search on "NVIDIA multi stage build" are:

• Smaller Docker Image using Multi-Stage Build
• Multi-stage Dockerfile with nvidia/cuda and Ubuntu NVIDIA/nvidia-docker#675
• A complete guide to building a Docker Image serving a Machine learning system in Production (section 3. Building a Docker image for any Python Project (GPU))

I'm not sure if mutli-stage builds help much though for Python projects. I think you would normally use the devel tagged nvidia/cuda images to build software and then once it is build copy it out in a secondary stage to a runtime tagged nvidia/cuda image.

[edmald]

Hi, @matthewfeickert what are the files that we will need to copy from devel to runtime to have a container with GPU computation capabilities and with a small size container?

[matthewfeickert]

@edmald for Python it isn't really practical / the devel is already as good as you're going to get without a huge amount of work (more than I have time for) as you need the relevant libraries available at runtime as well. I'd be happy to be proved wrong, but I don't have the bandwidth these days to delve into it more.