Jellyfin container works with NVIDIA GPU 1, but not GPU 0

[spongeboy03]

I'm having an issue when trying to change my jellyfin container from using my Tesla A2 to using my RTX 3090 (Both Ampere). Here is the output of nvidia-smi:

Code:

Fri Nov 21 20:02:46 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.105.08            Driver Version: 580.105.08    CUDA Version: 13.0    |
+-----------------------------------------+------------------------+----------------------+
| 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 3090        Off |  00000000:01:00.0 Off |                  N/A |
| 31%  43C    P8            17W /  350W |      15MiB /  24576MiB |      0%      Default |
|                                        |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|  1  NVIDIA A2                      Off |  00000000:02:00.0 Off |                    0 |
|  0%  38C    P8              6W /  60W |      14MiB /  15356MiB |      0%      Default |
|                                        |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU  GI  CI              PID  Type  Process name                        GPU Memory |
|        ID  ID                                                              Usage      |
|=========================================================================================|
|    0  N/A  N/A            1669      G  /usr/lib/xorg/Xorg                        4MiB |
|    1  N/A  N/A            1669      G  /usr/lib/xorg/Xorg                        4MiB |
+-----------------------------------------------------------------------------------------+

Here is my docker compose file. I tried adding the /dev/ mounts specified in the hardware acceleration page but it didn't fix it.

Code:

services:
  jellyfin:
    image: jellyfin/jellyfin
    container_name: jellyfin
    ports:
      - 8096:8096/tcp
      - 7359:7359/udp
    volumes:
      - ./config:/config
      - ./cache:/cache

      # Media
      - type: bind
        source: /mnt/omvpool/Movies
        target: /Movies

      # NVIDIA GPU device mounts
      - /dev/nvidia-caps:/dev/nvidia-caps
      - /dev/nvidia0:/dev/nvidia0
      - /dev/nvidiactl:/dev/nvidiactl
      - /dev/nvidia-modeset:/dev/nvidia-modeset
      - /dev/nvidia-uvm:/dev/nvidia-uvm
      - /dev/nvidia-uvm-tools:/dev/nvidia-uvm-tools

    restart: "unless-stopped"

    extra_hosts:
      - "host.docker.internal:host-gateway"

    runtime: nvidia

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["0"]
              capabilities: [gpu]

When I start the container with device_ids: ["1"], playback works flawlessly, but when I change it to use my 3090, it has a fatal error. I exec'ed into the container to see if jellyfin ffmpeg could use it for simple decodes, and I get this error:

[hevc @ 0x7f35548dd400] decoder->cvdl->cuvidGetDecoderCaps(&caps) failed -> CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected
[hevc @ 0x7f35548dd400] Failed setup for format cuda: hwaccel initialisation returned error.
[hevc_nvenc @ 0x7f35548dc600] OpenEncodeSessionEx failed: unsupported device (2): (no details)
[vost#0:0/hevc_nvenc @ 0x7f353fa7b100] Error while opening encoder - maybe incorrect parameters such as bit_rate, rate, width or height.
[vf#0:0 @ 0x7f3566281dc0] Error sending frames to consumers: Function not implemented
[vf#0:0 @ 0x7f3566281dc0] Task finished with error code: -38 (Function not implemented)
[vf#0:0 @ 0x7f3566281dc0] Terminating thread with return code -38 (Function not implemented)
[vost#0:0/hevc_nvenc @ 0x7f353fa7b100] Could not open encoder before EOF
[vost#0:0/hevc_nvenc @ 0x7f353fa7b100] Task finished with error code: -22 (Invalid argument)
[vost#0:0/hevc_nvenc @ 0x7f353fa7b100] Terminating thread with return code -22 (Invalid argument)
[out#0/matroska @ 0x7f3566281ac0] Nothing was written into output file, because at least one of its streams received no packets.
frame=    0 fps=0.0 q=0.0 Lsize=      0KiB time=N/A bitrate=N/A speed=N/A
Conversion failed!

Here is the command I used to test it. I downloaded jellyfin ffmpeg on the Debian host and ran the same command on the same file, and it was able to decode fine.

/usr/lib/jellyfin-ffmpeg/ffmpeg \
  -hwaccel cuda -hwaccel_output_format cuda \
  -init_hw_device cuda=cu:0 \
  -i /Movies/TEST_MOVIE.mkv \
  -c:v hevc_nvenc -c:a copy \
  -y "/cache/transcodes/test_output.mkv"

I am extremely curious as to why one gpu would work and not the other, when both are the same architecture, driver version, CUDA version, etc. The only difference between a working and non-working setup is that device_ids parameter.

[spongeboy03]

To add on, the 3090 works in my compute containers like ollama, vllm, etc. I shut those down while testing this.