missing steps in podman instructions when using nvidia GPU

[nalabrie]

Hello, I am running Fedora 40 Kinoite (Bazzite) with an NVIDIA GeForce RTX 4070 Ti (proprietary drivers, version 560.35.03). My aim was to install Jellyfin via a podman systemd container file and use GPU acceleration. I followed the instructions here but it did not work. After hours of debugging, I found a few issues with these instructions:

1. AddDevice=/dev/dri/:/dev/dri/ is not enough. I also needed to add AddDevice=nvidia.com/gpu=all and have the NVIDIA Container Toolkit installed on my system (podman specific docs here). I did not need to install this, on my machine it was already available so I cannot assist with proper install instructions. I am unsure if AddDevice=/dev/dri/:/dev/dri/ is still needed after this, I just left it there as well to be safe.
2. The instructions currently say SecurityLabelDisable=true is only needed for versions of container-selinux < 2.226. After making the changes mentioned above, it was still not working. I ran rpm -q container-selinux to see my version is container-selinux-2.232.1-1.fc40.noarch. I added SecurityLabelDisable=true anyway since the nvidia docs instruct to do so, and now everything is working.

Here is my full jellyfin.container file:

[Unit]
Description=jellyfin

[Container]
Image=docker.io/jellyfin/jellyfin:latest
AutoUpdate=registry
PublishPort=8096:8096/tcp
UserNS=keep-id
SecurityLabelDisable=true
AddDevice=/dev/dri/:/dev/dri/
AddDevice=nvidia.com/gpu=all
Volume=jellyfin-config:/config:Z
Volume=jellyfin-cache:/cache:Z
Volume=/mnt/elements/media:/media:Z

[Service]
# Inform systemd of additional exit status
SuccessExitStatus=0 143

[Install]
# Start by default on boot
WantedBy=default.target

In addition to verifying the GPU acceleration was working by transcoding some media successfully, I also ran podman exec -it systemd-jellyfin bash to get a bash shell inside the container and then ran nvidia-smi. Here is the output of that:

Mon Sep  2 16:43:10 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03              Driver Version: 560.35.03      CUDA Version: 12.6     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4070 Ti     On  |   00000000:2D:00.0  On |                  N/A |
|  0%   42C    P8             11W /  285W |     649MiB /  12282MiB |     11%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+

If the container could not access the GPU then nvidia-smi would return command not found. I can also see the GPU load increasing when using a system monitor and playing media that needs transcoded.

[nalabrie]

I'm happy to make a PR but wanted to post this as an issue first to see if anyone else can confirm if these changes are needed or if I'm missing/misunderstanding anything.

[felix920506]

If you are unsure please ask in our matrix/discord channels. https://jellyfin.org/contact