Fatal Play back Error on some media

[SleepyPillow-bot]

Hey all,
I’m trying to get Nvidia GPU transcoding working in Jellyfin (running as a Docker container on Ubuntu 24.04, with an RTX 2080 Ti). I keep hitting what looks like a CUDA permissions or access problem, and nothing I do seems to fix it.
What’s happening:

• Whenever Jellyfin tries to transcode with NVENC/NVDEC, the logs show:
  

[AVHWDeviceContext @ ...] Cannot load libcuda.so.1
Could not dynamically load CUDA
Failed to set value 'cuda=cu:0' for option 'init_hw_device': Operation not permitted
Error parsing global options: Operation not permitted

Direct play and direct stream work. Only hardware transcoding is broken.
What I’ve tried and checked so far:

• On the host, nvidia-smiworks — 580.65.06.
  
• Installed nvidia-cuda-toolkit so CUDA libraries are present. Confirmed /usr/lib/x86_64-linux-gnu/libcuda.so.1 points to the correct file.
  
• All /dev/nvidia* devices have group video, and my host-level checks look good.
  
• I’m using the official Jellyfin Docker image and keeping up to date.
  

Inside the Jellyfin UI:

• All hardware codecs for my GPU are enabled (no AV1, obviously for 2080 Ti).
  
• Hardware encoding and "HEVC encoding" switched on.
  

Testing:

• Running Jellyfin’s exact ffmpeg command inside the container gives the same CUDA load/permission error.
  

At this point:
I’ve rebooted, regeared, checked and re-checked guides and docs for Jellyfin+Docker+Nvidia, but it's still not working. The container sees the GPU (nvidia-smi works in the container with --runtime=nvidia), but every attempt to use NVENC/NVDEC fails in ffmpeg (from either Jellyfin or the console).
Really appreciate any advice—this is driving me nuts!
Thanks!