nvidia-uvm kernel module build fails on debian 13 kernel 6.12

[sweet robin]

I tried to install nvidia drivers on my debian 13 with kernel 6.12 for my rtx 5060 ti but it fails.

I know that per this wiki: https://wiki.debian.org/NvidiaGraphicsDrivers I cannot use prepackaged drivers so I downloaded all recent versions, namely:

• 580.105.08
• 580.126.09
• 590.48.01

They all more or less fail at building nvidia-uvm kernel module with these undefined symbols while building modules.

It tells me to disable UVM but that would be a problem since I plan to use this nvidia gpu in my kubernetes cluster for av1 encoding and inference

ERROR: An error occurred while performing the step: "Checking to see whether the nvidia-uvm kernel module was successfully built". See /var/log/nvidia-installer.log for details.
-> The command `cd kernel-open; /usr/bin/make -k -j32  NV_EXCLUDE_KERNEL_MODULES="" SYSSRC="/lib/modules/6.12.43+deb13-amd64/source" SYSOUT="/lib/modules/6.12.43+deb13-amd64/build" NV_KERNEL_MODULES="nvidia-uvm"` failed with the following output:

make[1]: Entering directory '/usr/src/linux-headers-6.12.43+deb13-common'
make[2]: Entering directory '/usr/src/linux-headers-6.12.43+deb13-amd64'
  MODPOST /tmp/makeself.giyDQPjm/NVIDIA-Linux-x86_64-580.105.08/kernel-open/Module.symvers
ERROR: modpost: "nvUvmInterfaceGetEccInfo" [/tmp/makeself.giyDQPjm/NVIDIA-Linux-x86_64-580.105.08/kernel-open/nvidia-uvm.ko] undefined!
ERROR: modpost: "nvUvmInterfaceP2pObjectCreate" [/tmp/makeself.giyDQPjm/NVIDIA-Linux-x86_64-580.105.08/kernel-open/nvidia-uvm.ko] undefined!
ERROR: modpost: "nvUvmInterfaceGetChannelResourcePtes" [/tmp/makeself.giyDQPjm/NVIDIA-Linux-x86_64-580.105.08/kernel-open/nvidia-uvm.ko] undefined!
WARNING: modpost: suppressed 66 unresolved symbol warnings because there were too many)
make: *** [Makefile:138: modules] Error 2

[woeful perch]

I cannot use prepackaged drivers
May I ask why? You've pointed to the Wiki page on installing the driver, but I don't understand why you can't use the prebuilt versions
You can also follow NVIDIA's official guide on installing the driver (which is referenced in the Debian Wiki as well)

[sweet robin]

it is marked on the nvidia wiki that no prepackaged versions support blackwell gpus for nvidia.

However despite all that I tried them all, I tried everything, the link you gave is the first thing I tried and it failed as well.

I even tried building modules manually from open-gpu-kernel-modules github and ended up opening a ticket because it still didn't work:

https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1006

[sweet robin]

would you happen to know if there are quirks related to this specific setup?

[plush laurel]

NOTICE: Very new hardware (e.g. RTX 50 series) may have spotty support on older or LTS releases/builds of Linux distributions. Double check your distro if you are not comfortable with manually tinkering.

It is assumed you know how to create and boot from a Linux distro's installation media (including but not limited to downloading ISOs, burning/imaging ISOs to a DVD/USB drive, and disabling Secure Boot temporarily)

• Unless you know what you are doing, and unlike on Windows, do NOT use the binary packages downloaded from the NVIDIA website. Check your distro's docs/wiki on how to install NVIDIA drivers correctly.
• Change your communication protocol to Wayland (for performance) if your distro defaults to the old X11 (for compatibility). Most newer distros and NVIDIA drivers should just work, but a handful of distros still do not have a fully working DE with Wayland or the driver it ships with.
• If the installer or your installed OS boots into a black screen, try using an alternative driver mode, graphics mode, or setup mode. Some distros do not have functioning in-box drivers for newer NVIDIA graphics cards.
• If your laptop relies on hybrid graphics, make sure that your drivers, desktop environment, and communication protocol are up to date.
• DLSS support requires an up-to-date version of Proton; DLSS Frame Generation may not function correctly.
• Usually, GPU offloading will be handled automatically by DXVK, VKD3D-Proton or Winevulkan for non-native games, and PRIME for native games.
• PRIME is supported by the closed NVIDIA driver (for most current GPUs) and the Nouveau driver (for legacy GPUs).
• Do not use older, unmaintained software such as nvidia-xrun or Bumblebee.

[sweet robin]

I don’t need gaming only cuda. I tried different solutions beginning with debian information on how to install the drivers. It doesn’t change much

I don’t have nouveau I don’t use proton/wine. wayland or xorg are not relevant since this isn’t used to display anything. Not even plugged in

[woeful perch]

The open kernel modules support Blackwell hardware. I assume you've tried installing those specifically? (Nvidia's Wiki mentions both options)
I am not a Debian user, but reading the Wiki, I would assume the command you need is

sudo apt -V install nvidia-driver-cuda nvidia-kernel-open-dkms

Have you tried installing those specific packages already? If so, how did that command fail?

[sweet robin]

Yes I tried the open kernels. Debian docs are pretty clear about the fact that there is no proprietary packaged drivers. These are the ones I fail to install

[woeful perch]

How exactly do they fail to install? Could you share any error messages and/or the output of the install command?

[sweet robin]

I linked an issue opened on github of open kernel modules. The error logs are there. Discord doesn’t allow long messages. And I followed the readme of that repo

[sweet robin]

@woeful perch I can share the logs here as well fi you want, here is a snippet:

The exported twice modpost error appears thousands of times