Issue Installing Cuda-drivers 440.64.00 Keeps trying to use 450

Graphics / Linux Linux
1

OS: CentOS

we have hundreds of servers to manage Software versions using ansible.
two of our servers have both 440.64.00 packages without issue
using
nvidia-drivers-branch-440
which installed 440.64.00 fine

the third today:
installing drivers-branch 440.64 worked fine

on upon installing cuda-drivers-440.64.00
i get this error
NVIDIA: No kernel module package kmod-nvidia-latest for kernel-3.10.0-1062.4.1.el7.x86_64 and 3:nvidia-driver-latest-450.36.06-1.el7.x86_64 found. Ignoring the new kernel
Error: Package: 3:nvidia-driver-branch-440-440.64.00-1.el7.x86_64 (@nvidia_cuda_compute)
Requires: nvidia-kmod = 3:440.64.00
Removing: 3:kmod-nvidia-latest-dkms-440.64.00-1.el7.x86_64 (@nvidia_cuda_compute)
nvidia-kmod = 3:440.64.00
Updated By: 3:kmod-nvidia-latest-dkms-450.36.06-1.el7.x86_64 (nvidia_cuda_compute)
nvidia-kmod = 3:450.36.06

  1. it appears to be ignoring that want 440.64.00

this was fine on Friday as i installed it on 2 other servers (ansible below)

- name: install package requirements

package:
name:
- kernel-headers
- kernel-devel
- “yum-plugin-nvidia-{{ yum_nvidia_plugin_version }}”
- “nvidia-driver-branch-{{ nvidia_drivers_branch }}”
- “cuda-drivers-{{ nvidia_cuda_driver_version }}” # issue on this package
- “nvidia-container-toolkit-{{ container_toolkit_version }}”
state: present

2

driver 450 comes with cuda 11(rc). I guess you have initially installed cuda by using the “cuda” metapackage which means that everything will be upgraded automatically.
Cuda 11 is now in repo:
http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/

3

@generix, Your point makes sense. however this is a compleltly new install
as you see from the code No where relates to 450 so not sure however it interfered with 440. so this should not happen

to clarify ive only installed
nvidia-drivers-branch-440
cuda-drivers-440 (which is where the error happen)

so if i install cuda 10.2 meta then this issue could disappear?

4

I had gpu device breakage over the weekend with docker containers and new installation of select packages from cuda-10-2. When it broke I went to install the drivers into the container so I could run nvidia-smi, but the driver installation failed (as you saw in my other post). Similar to your post, I suspect it is installing newer components from the 450 driver and breaking things, but I can’t even get the driver installed on a clean centos:centos7 container at the moment.

5

I’ve taken a look at the packages provided by nvidia-driver-branch-(id) and this should pull in nvidia-driver-branch-440-cuda-libs-440 which contains libcuda.so (the cuda driver). So in my understanding, when using the branched driver packages, installing cuda-drivers is not useful for this just being a meta package for handling driver upgrades. The documentation of those branched driver is horribly meager, was source of confusion before.
Of course, you will then have to install the matching cuda-toolkit-10-2 meta package and hope this doesn’t rely on cuda-drivers meta.

6

thanks @generix makes sense will ignore & test without cuda-drivers. entirely first so far -branch has been sufficent