Question : Could we use `conda install cudatoolkit` as a replacement? about gpu-jupyter HOT 4 CLOSED

Hi,

Thanks for the quick answer. I think I was a bit unclear (sorry for that).

We are running docker and we are basing our work on top of docker-stacks (datascience notebook to be exact)
Our current approach is to (code here if interested https://git.cs.kau.se/jonakarl/jupyterhub) :

1. Create a custom container based on the datascience-notebook
2. Patch the build scripts of nvidia-container to use our image (in step 1) instead of ubuntu20.04 (datasceince notebook also uses ubuntu20.04)

While it works step 2 is hacky at best and not very clean.
Therefore we are evaluating to instead base our custom image on top of your image or alternatively install cudatoolkit and add the ENV variables + possible xtras needed to make it run as a GPU container.
I guess the "possible xtras" is the key, how much "magic" do you or nvidia do to make it work :-)

from gpu-jupyter.

Thanks for the valuable input.

Closing this now as I think I have all info I need, will play a little with the options.

from gpu-jupyter.

Good question!

An installation on top of Anaconda is also a good choice and will provide the same or at least similar packages.
From my point of view, these are the main advantages of using GPU-Jupyter (or a similar option):

• The toolstack for data science, analytics and scientific computing is based on jupyter/docker-stacks which is very robust and provides some additional selected packages. One can also change the underlying stack, e.g., the full option here is based on the docker-stack's datascience-notebook that provides the Julia language and R for collaborative work.
• Dockerhub stores pre-built images with already installed GPU-libraries. Pulling and running Images from Dockerhub is usually faster and simpler than manually installing them.
• Docker comes with additional advantages: Many prefer solutions in Docker as one can easily define and modify the installation of a setup. This is especially interesting for the scientific community, because it makes it easy to share a specific setup on which experiments were executed.

from gpu-jupyter.

While it works step 2 is hacky at best and not very clean.

That was also the reason for me to build the docker-stack on top of nvidia-supported image. There are multiple ways to get to a good solution. This repo provides one that suits for many common applications and is intended for an easy setup and customization.

Therefore we are evaluating to instead base our custom image on top of your image or alternatively install cudatoolkit and add the ENV variables + possible xtras needed to make it run as a GPU container.
I guess the "possible xtras" is the key, how much "magic" do you or nvidia do to make it work :-)

I think a proper option would be to just fork this repo and do all necessary changes. Custom ENV variables can be placed, e.g., into the docker-compose.yml. Possible xtra installations should be placed into src/Dockerfile.usefulpackages. Make also to remove unused packages from this file if wanted.

The >best< setup depends a lot on the actual use case. Therefore, the src/Dockerfile.usefulpackages is separated to enable customization for a specific use case, based on a robust GPU and data science toolstack.

from gpu-jupyter.