Bitbucket runner and NVIDIA runtime Edited

[Piotr Nowak]

Hi everyone,

Short background: I am developing Python project which requires GPU access (CUDA).

I have experimented with bitbucket pipelines runner on the workstation with GPUs but unfortunately I cannot make it to work correctly. I am able to build a container with all the required CUDA dependencies etc inside the step script, but afterwards it cannot be run as it is essential to use nvidia-docker runtime https://github.com/NVIDIA/nvidia-docker.

I am passing `-v /var/run/docker.sock:/var/run/docker.sock` from host to the runner, but it does not look like it is passed through to the container running given step.

The docker service which can be added to each step does not help as it seems to use `docker-public.packages.atlassian.com/sox/atlassian/bitbucket-pipelines-docker-daemon:v19-prod-stable` image which does not use nvidia-docker runtime.

I have a question if and how could it be done - I have spent already several hours trying to resolve that (in simpler or more complex ways), but I am running out of feasible ideas now.

Thanks in advance for any advice.