"Cannot load libcuda.so.1" unable to do HW transcoding

Plex Media Server
,
1

Environment

  • Plex Media Server: 1.43.2.10687

  • Also tested: 1.43.1.10611 and 1.42.2.10156

  • Image tested:

    • linuxserver/plex
    • plexinc/pms-docker

  • Host OS: Debian Trixie

  • Docker: Docker Compose

  • GPU: NVIDIA RTX 3060

  • Driver: 550.163.01

  • CUDA: 12.4

  • NVIDIA Container Toolkit: 1.19.1

  • Plex Pass: Lifetime

    Every transcode attempt produces:

    Cannot load libcuda.so.1
Could not dynamically load CUDA
opening hw device failed - Operation not permitted
Could not create hardware context

    Plex then falls back to software transcoding.

Inside the running Plex container:

libcuda.so.1 exists

ldconfig -p | grep libcuda

libcuda.so.1 => /usr/lib/x86_64-linux-gnu/nvidia/current/libcuda.so.1

Plex Transcoder reports:

Hardware acceleration methods:

cuda
vaapi
drm
opencl

and

h264_nvenc
hevc_nvenc

are listed as available encoders.

I installed Jellyfin on the same Docker host, using:

  • same NVIDIA runtime
  • same RTX 3060
  • same Docker
  • same NFS media mount

Jellyfin successfully starts GPU transcoding.

I’m getting the feeling I’m missing something elementary, but I’m buggered If I know what it is.
Works for Jellyfin just fine, but Plex just refuses to use the same hardware on the same docker host.

2

something specific to debian trixie ffmpeg and the supported nvidia driver for DT.

Rebuild the VM as ubuntu with the exact same hardware and config and it just works immediately