CUDA not permitted

[natzilla]

@nebulon I don't mind at all, I will make some edits to my ACL's rules to allow ssh and send you the details in the email.

[natzilla]

@nebulon
I have sent the email titled "Jellyfin test instance for CUDA" all the details for access are in it.

[nebulon]

After some investigation, it turns out that one has to use a custom docker version provided by nvidia to add support for this. https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#getting-started and https://github.com/NVIDIA/nvidia-docker

So this means that we have to somehow add support on the Cloudron platform side as well as make sure things are working in nvidia and then also non-nvidia case, given that docker parts have to be compatible with whatever the server currently provides hardware wise.

So this is a bit out of scope for the coming release but hopefully we can get this supported afterwards.

[natzilla]

@nebulon

I'm surprised I missed that, but I'm also surprised Nvidia had to make this difficult. Going forward I'm willing to provide you the access you need for testing implementation whenever that can happen. I'll have to live with none hardware acceleration for now.

[robi]

@nebulon this appears to use a docker configuration in privileged mode to get access to the hardware device, which goes against the cloudron use case and security posture.

A better approach would be to use a different runc such as sysbox-runc that can solve that issue without requiring privileged mode.

[natzilla]

@nebulon

Hi, it's been a little while since this was last looked into. With the suggestion from robi do you think this is something we can look at doing soon? I am still willing to provide access for testing CUDA.

[AartJansen]

@nebulon it looks like that nvidia-docker has been superseded by nvidia container toolkit
https://github.com/NVIDIA/nvidia-container-toolkit
Does that make the issue go away or change anything at all ?
I'm also happy to allow you access to my cloudron server with nvidia GPU to try out.

I've got the nvidia drivers installed, and nvidia container toolkit installed also, but...

In this guide
https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
I have stopped at running the configuration command

sudo nvidia-ctk runtime configure --runtime=docker

because the file /etc/docker/daemon.json doesn't exist... (and I can't find it anywhere else either) which leads me to think it's not going to work.
Also theres nothing explaining what effect it will have on the existing docker containers, I'm assuming it will add access to the GPU to all containers.

[robw]

The file is created if it doesn't exist already, I think that command is safe to run:

[AartJansen]

Did it make any difference?
Assuming its functioning I would think the docker start would still have to be modified.
ie docker run --gpu all <container> where ever that is.

[AartJansen]

Ended up removing jellyfin from cloudron, its basically useless without hardware transcoding, imo.

I followed these instructions https://www.geeksforgeeks.org/how-to-install-jellyfin-media-server-on-ubuntu-22-04/ to install it in ubuntu (i know its not a supported config). And streaming with nvidia support makes it work like a different product. I don't have fancy DNS or SSL certs ...but it works

[robi]

@AartJansen I still think more progress could be made if a different docker runtime would be used as mentioned in #10