Add page on creating a RAPIDS container from an Ubuntu base image

[jayavenkatesh19]

Follows #582

Now that we have a page on the deployment docs describing how to build a RAPIDS container from the cuda containers as the base image, it would be good to have a page that installs from a base image such as Ubuntu. This will be very useful for RAPIDS users who do not have the flexibility to choose their base images and would want to install RAPIDS on top of their existing Docker builds.

[ncclementi]

Putting this here, as it will help creating this content https://gist.github.com/jacobtomlinson/30ad4126d52e62bdd13f3a05b4d8408c

[jayavenkatesh19]

Looking into this more closely, there are a few considerations to be made in designing a one-fits-all solution for this:

1. Recommending the conda route towards installing RAPIDS in a container is pretty straightforward. The CUDA base images like the one here for CUDA 13 on Ubuntu 24.04 install the cudart and compat packages in which the versions are hardcoded based on the CUDA version and the Linux distro of the image. This fits in perfectly with the conda template Dockerfile in our Custom RAPIDS Docker page, and I was able to start from an Ubuntu image and get a working RAPIDS installation, just by installing these two packages as described in the Dockerfile above.

2. Things get a little tricky on pip on the other hand. pip installations would require us to start from a CUDA runtime image (because of the well documented cupy not being able to find libcudart issue). Looking at the Dockerfile for the runtime images, like this one here for CUDA 13 on Ubuntu 24.04, we have a lot of packages that need to be installed whose versions are hardcoded in the Dockerfile itself. So while we can have some information about the Dockerfile and its components, having a template on the Deployment docs would be hard to create due to sheer number of versioning differences in these packages for different CUDA and Linux variants.

One hacky method around this I have been experimenting with is to copy the CUDA installation from the base/runtime image matching the user's environment to replicate the CUDA libraries and symlinks on their image. Something like this:

FROM nvidia/cuda:12.9.1-runtime-ubuntu24.04 AS cuda-source
FROM ubuntu:24.04
COPY --from=cuda-source /usr/local/cuda-12.9 /usr/local/cuda-12.9
COPY --from=cuda-source /usr/lib/x86_64-linux-gnu/libnccl* /usr/lib/x86_64-linux-gnu/

This works perfectly for both the base and runtime images, and I was able to get both conda and pip containers of RAPIDS to work starting from an Ubuntu image (admittedly only tested this with some cudf dataframe and series creation). But this might not be a solution we would be able to recommend because this essentially orphans these libraries from apt (or any other package manager), and possible licensing issues.

All things considered, I think the best way to represent this knowledge on the Deployment docs is to have a page explaining the various components of the CUDA Dockerfiles themselves with an example showing how to install CUDA components on an Ubuntu image. The alternative would be copying the entire CUDA installation from the CUDA images onto the user's Docker image.

[jacobtomlinson]

have a page explaining the various components of the CUDA Dockerfiles themselves with an example showing how to install CUDA components on an Ubuntu image

Yes this sounds great. Some kind of diagram of components would be awesome too.