docker container does not run on GPU

[DiTo97]

Describe the bug
we have built a docker container pulling from the official NVIDA repository hub:

docker pull nvidia/cuda:11.8.0-devel-ubuntu22.04

and ran the procedure to install rerun following your documentation.

rerun works (ran a few examples), but detects only the CPU as rasterizing device. The container easily sees the GPU (NVIDIA A10) well, as we have other workloads in that container that run on the GPU with no issues at all when invoked.

we have also run rerun in the host machine (AWS EC2) outside the container and it works on the GPU just fine.

[jleibs]

Can you see if the instructions from:

• Add troubleshooting or how-to guide related to running Rerun in a Docker container #6835

address your issue?

[DiTo97]

@jleibs, I tried, but could not see any real difference.

[DiTo97]

@jleibs, what's the roadmap for official docker support?

[jleibs]

@jleibs, what's the roadmap for official docker support?

This is not an item on our roadmap at the moment.

"Docker", unfortunately creates a combinatorial explosion of complex host/container environments. In my experience 95% of the issues we have seen have been related to configuration of the Docker or host environment rather than an actual issue with the Rerun view.

We do know that Docker deployments can be made to work.

The instructions from #6835 continue to work for me.

If it is helpful to your debugging, I am running on Arch Linux, using wayland via hyprland, and an RTX 4070

>  nvidia-smi 
Wed Sep  4 17:22:59 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03              Driver Version: 560.35.03      CUDA Version: 12.6     |
|-----------------------------------------+------------------------+----------------------+

My docker, and nvidia-container package versions are:

docker 1:27.2.0-1
libnvidia-container 1.16.1-1

If you have followed the linked instructions verbatim, I would deduce the issue may have something to do with the host nvidia driver version, or even NVIDIA A10 support within driver or container runtime support version.

we have other workloads in that container that run on the GPU

Can you confirm that these are vulkan applications that render to a display? My go-to test application is typicall vkcube. If this finds your GPU but Rerun doesn't, that might help narrow in on the nature of the problem.

All that said, if you are unable to get Rerun to work inside the docker container, I would recommend using the viewer on one of the many supported native environments, or via the web, and just remotely access the data from the docker container without using that environment to do the rendering itself.