seeking a way to install newest nvidia driver while transcoding on Debian system

[hypokamp]

Hi there,

I host a Jellyfin server installed on a LXC container on a Debian 12 system on my Proxmox home server. I use my Nvidia GPU for transcoding and I followed the official tutorial for the installation https://jellyfin.org/docs/general/admini...on/nvidia/
Everything works fine.

My issue is that for Nvidia GPU transcoding to work with Jellyfin, I had to install the 2 packages libnvcuvid1 libnvidia-encode1 from the Debian repository and these packages are old. Installing them forces me to use old Nvidia drivers. This is for me an issue because I need newer drivers to install CUDA drivers and use my GPU for other purpose such as AI training.

My question is this : is there a way to transcode in Jellyfin on a Debian system without requiring these 2 old packages ?
* Can I for instance use Nvidia Video codec SDK instead of them ?
* Can I recompile these 2 packages with newest Nvidia drivers and if so, can you point me to some resources about how to recompile them ?
* Do I need to recompile ffmpeg7 with the newest Nvidia driver ?
* Is there any other way ?

[TheDreadPirate]

If you used a VM instead of a LXC, would you still have to install the old Nvidia drivers on the host? Could you still share a GPU between a LXC and a VM?

AFAICT, using a VM would allow you to fully isolate drivers on that VM. But I can't get a clear answer on whether that GPU is still sharable with older VMs and LXCs.

[hypokamp]

If you used a VM instead of a LXC, would you still have to install the old Nvidia drivers on the host?  Could you still share a GPU between a LXC and a VM?

AFAICT, using a VM would allow you to fully isolate drivers on that VM.  But I can't get a clear answer on whether that GPU is still sharable with older VMs and LXCs.

Thanks for your reply 🙂

I use LXC containers because I can share my GPU resources between containers running concurrently. That imposes to install the same driver on the host and the containers.

I could use a VM and install specific drivers for it. However I wouldn't be able to share the GPU with other VM/containers that way, so that's not what I want to do.

My goal is to run a Jellyfin server alongside a custom LLM for example.

Also, in the VM setup, switching between GPU-in-VM to GPU-to-LXC is very complicated and would require to reboot the server each time, so that's a big no for me.