docker build guide doesn't work

[jonthegeek]

I'm trying to build a torch-friendly environment in docker using docker/build_env_guide.md, but I'm hitting errors. I'll specify them below, but it looks like the guide was updated (it mentions cuda-113 and cuda-102 directories) but the accompanying files were not (the directories are cuda-111 and cuda-101).

The first thing I had to change was the FROM nvidia/cuda line in the Dockerfile. FROM nvidia/cuda:11.1-cudnn8-devel-ubuntu18.04 doesn't exist as far as I can find. I updated that to 11.1.1, and then after that failed also tried ubuntu20.04, although I don't think that's actually the problem.

build_env.sh works (with the updated FROM), but docker-compose up -d gives this error:

ERROR: for rstudio_docker  Cannot create container for service rstudio_docker: Unknown runtime specified nvidia

I assume that's an issue with line 7 of docker-compose.yml (runtime: nvidia), but I don't know enough about docker or docker-compose to grok what needs to change.

I'm running this on Windows in case that matters.

[dfalbel]

@y-vectorfield Do you have any ideas how to fix this?

[y-vectorfield]

@jonthegeek , @dfalbel
Thank you very much for your message.
Actually, NVIDIA GPU runtime for Docker has been updated recently.
Therefore, this Dockerfile need to modify.
I will modify this and will submit a PR.

[y-vectorfield]

@dfalbel
Is this issue active?
If this is active, I will deal with this problem.
Actually, I tried to modify this previously. However, I could not solve this...
I will retry this.
I will create a new builder using Docker.

[dfalbel]

Hi @y-vectorfield,

I still didn't have time to work on a new guide. I have worked on different approach here though: https://github.com/mlverse/docker/tree/main/torch

[y-vectorfield]

Hello, @dfalbel
OK, I will help you to create new build tools.

[y-vectorfield]

@jonthegeek, @dfalbel, I'll help to enhance materials for docker build.