Jellyfin 'fatal player error' after an NVIDIA update: the driver/library mismatch

Some videos in Jellyfin started failing with a generic “fatal player error”. Others played fine. The pattern: everything that needed transcoding failed, like HEVC files or files with E-AC3 audio on a client that can’t play it. Files that direct-played worked.

It looked like a codec problem. It was actually the GPU driver.

The diagnosis

$ nvidia-smi
Failed to initialize NVML: Driver/library version mismatch

That one message explains everything:

• Ubuntu’s unattended-upgrades had updated the NVIDIA driver packages in the background, for example from 580.159 to 580.173.
• The kernel module that was actually loaded stayed on the old version, because that only changes on a reboot.
• My server had been up for 87 days, so it ran with mismatched user-space libraries and kernel module for weeks.

Any container that started using the GPU after the update failed to initialise it. Jellyfin’s hardware transcoding failed, and the player only showed “fatal player error”.

You can confirm the mismatch by comparing the two versions:

cat /proc/driver/nvidia/version      # what's loaded right now
dpkg -l | grep nvidia-driver          # what's installed

The fix

A reboot. The new module loads, nvidia-smi works, and transcoding comes back.

Preventing it: reboot only when needed

I didn’t want to reboot every night, and I didn’t want to remember to check. So a small script runs every day at 05:00 from cron and reboots only if there’s a reason:

#!/bin/bash
PATH=/usr/sbin:/usr/bin:/sbin:/bin
reason=""
[ -f /var/run/reboot-required ] && reason="pending kernel/package update"
if nvidia-smi 2>&1 | grep -qi "version mismatch"; then
    reason="${reason:+$reason; }NVIDIA driver/library mismatch"
fi
if [ -z "$reason" ]; then
    logger -t scheduled-reboot "No reboot needed"
    exit 0
fi
logger -t scheduled-reboot "Rebooting, reason: $reason"
sync
systemctl reboot

And in /etc/cron.d/scheduled-reboot:

0 5 * * * root /path/to/scheduled-reboot.sh

On quiet days it does nothing. When updates land, roughly twice a month, the server reboots at 5 a.m. while everyone’s asleep, and a GPU mismatch never lasts more than a day.

One important detail: the script does not run docker stop before rebooting. I learned that the hard way.

Alternatives

• apt-mark hold the NVIDIA packages and update them by hand. Fine if you remember to.
• Enable Unattended-Upgrade::Automatic-Reboot "true" in unattended-upgrades. It’s simpler, but it doesn’t catch the NVIDIA-specific case.

Either way: if GPU features in Docker suddenly stop working after an update, run nvidia-smi on the host first.

Quick checklist when GPU transcoding breaks

1. nvidia-smi on the host. Mismatch error → reboot.
2. docker exec jellyfin nvidia-smi. If the host works but the container doesn’t, the container’s GPU runtime config is wrong (runtime: nvidia or a deploy.resources device reservation).
3. Jellyfin → Dashboard → Playback: is NVENC still selected, and are the codecs your card supports ticked? A GTX 1050 Ti can decode HEVC 8/10-bit but not AV1.
4. Check Jellyfin’s transcode log for the actual FFmpeg error. It’s far more specific than “fatal player error”.