↓ Skip to main content

Nvidia GTX driver shenanigans on Trixie and forced Frankenization

·363 words·2 mins·
Author
Tuco Imad

The NVIDIA Driver Survival Story on Debian Trixie
#

So, I decided to run some advanced computations on my Debian machine. And to do such I needed better drivers. Simple enough, right?

Phase 1: Fresh install - from Nouveau free stock to Nvidia Proprietary
#

Installed. It booted. It rendered a desktop. It did GPU things.

No CUDA, no real acceleration, but at least nothing was actively breaking. A stable baseline to destroy later. Then I added the keyring, hit apt update, got the packages, installed nvidia-driver metapackage which promptly pulled latest and… Hello black screen.

Phase 2: Fixing the Black screen, realizing NVIDIA 590+ dropped GTX
#

So, upon my realization that the 590 dropped GTX and only supports RTX (was pretty evident from nice error logs post-install, as well as the black screen, lol. I switched to the official Debian 12 CUDA repository:

https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/

Purged with apt all Nvidia-related things and PINNED all higher packages to not install. Then, installed NVIDIA driver branch 580.x.

At this point:

  • CUDA works
  • GPU acceleration works
  • system is actually usable
  • Wayland mostly behaves

Solid.

What was the most annoying was directly trying to downgrade from 590 back to 580, because I got:

  • dependency conflicts everywhere
  • APT refusing partial rollback
  • meta-packages dragging wrong versions back in
  • mixed driver states in /usr/lib and kernel modules

The actual fix was not “downgrade nicely”.

It was:

  • purge broken 590 packages
  • remove conflicting CUDA driver meta packages
  • reinstall clean 580 branch from Debian 12 CUDA repo
  • ensure everything is pinned to 580 series

Once done:

  • system stabilised again
  • CUDA returned
  • GPU stack became consistent

Phase 3: Wayland breakage episode
#

After stabilization, I spent ~2 hours debugging Wayland issues.

Symptoms:

  • session starts then crashes or hangs
  • inconsistent rendering behavior
  • GPU initialization not matching expectations

Root cause turned out to be:

  • missing / misconfigured GDM / session backend expectations
  • GPU modeset flags not consistently applied
  • NVIDIA DRM configuration not aligned with Wayland requirements

Required kernel + boot configuration
#

This is what ended up mattering:

/etc/default/grub
#

GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on modprobe.blacklist=i915 kvm.ignore_msrs=1 vfio-pci.ids=8086:1912 snd-intel-dspcfg.dsp_driver=1 snd_hda_core.gpu_bind=0 nvidia-drm.modeset=1 nvidia-drm.fbdev=1"

After that, the system just worked ever since and pinning solved the upgrading issues! Yay!