Nvidia production driver not found on debian13 repo

Graphics / Linux Linux
1

You’re providing debian packages of the nvidia driver, with instruction given at

The repos are located at Index of /compute/cuda/repos
I can see you’re supporting debian13, debian12, debian11, debian10. That’s the current stable Debian release (debian13 = trixie), and a few older releases.

That’s great to do this. But the problem is that you have a discrepancy with the nvidia production driver release. The current nvidia production version is 580.142 (currently listed at Unix Drivers | NVIDIA )

But nvidia’s debian13 repo only provides nvidia 590 and 595.

This means that it is not possible to install the stable nvidia driver on a debian stable system, which is kind of disastrous.

It’s disastrous in particular for older GPU cards such as Maxwell which have been dropped from from the 590 driver.

Can you please make the production nvidia driver (580) available on the debian stable repo?

I’d additionally suggest also making all the currently supported drivers available, which would mean the legacy 570 driver as well.

p.s. I meant to add another point of context: release 580 is required for building the nvidia kernel modules for linux kernel 4.9, which is why the old nvidia 550 versions provide by debian are no longer sufficient.

2

The repos you linked to are datacenter repos: the status of DC drivers is different than that of desktop graphic drivers as they are focused on compute workloads.

Unfortunately v580 is unlikely to be provided for D-13 because of the reasons I mentioned here:

If you need v580 to for older cards, a hacky workaround is to use the packages from the D-12 repo as described on the Debian wiki, DC section, D-13 subsection: NvidiaGraphicsDrivers - Debian Wiki

3

Ah I see. The production release is blocked by glibc-2.41+, meaning cuda v580 can’t build easily for debian13 (I wonder if the patch given at "error: exception specification is incompatible" for cospi/sinpi/cospif/sinpif with glibc-2.41 - #3 by stefantalpalaru is all that it needs). Thanks for pointing out the context for the missing production driver.

Thanks also for pointing out the workaround with the debian12 repo. I can confirm it does still work (and even with debian unstable). A little annoying to lose the security of the apt key, but the required allow-insecure=yes option can be easily enough configured into the debian12 repo definition in /etc/apt/sources.list, so the standard debian repo continues using its secure signatures,

deb [signed-by=/usr/share/keyrings/cuda-archive-keyring.gpg allow-insecure=yes] https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/ /
1 Like
4

I will add this method to the wiki: I wasn’t aware that it is possible to put so much magic into square brackets there ;-]

1 Like
5

I used to build a lot of CUDA apps on Debian-13 with such a patched header before repos dedicated to D-13 were created and have never experienced any problems (I mean “additional problems”, not present on other officially supported configs).