Installing nvidia-cuda-toolkit removes nvidia-driver-510

[Gon Solo]

I tried to install the CUDA toolkit.
When doing a "sudo apt install nvidia-cuda-toolkit" apt tries to remove my driver "nvidia-driver-510". Also the cuda toolkit is 11.3 whereas 11.6 is available.

I think the two packages should be in sync.

ProblemType: Bug
DistroRelease: Ubuntu 21.10
Package: ubiquity 21.10.10
ProcVersionSignature: Ubuntu 5.13.0-30.33-generic 5.13.19
Uname: Linux 5.13.0-30-generic x86_64
NonfreeKernelModules: zfs zunicode zavl icp zcommon znvpair nvidia_modeset nvidia
ApportVersion: 2.20.11-0ubuntu71
Architecture: amd64
CasperMD5CheckResult: unknown
CurrentDesktop: ubuntu:GNOME
Date: Mon Mar 7 17:23:07 2022
InstallCmdLine: BOOT_IMAGE=/boot/vmlinuz-5.13.0-30-generic root=UUID=5ad16aab-e3c0-47b9-a0fd-f06aaf640323 ro quiet splash vt.handoff=7
SourcePackage: ubiquity
Symptom: installation
UpgradeStatus: No upgrade log present (probably fresh install)

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Russell Morris (6-u3untu-h)]

Hi,

The same issue here - it wants to force me back to nvidia-driver-470. Is there a way around this?

Thanks!

[Joseph Yasi (joe-yasi)]

This is a debian packaging issue.
rules.defs in the Debian source package has different rules for Debian and Ubuntu for libnvidia-ml1 and libcuda1. From the jammy package of 11.5.1-1ubuntu1:
-----snip-----
package_libcuda1.Debian += libcuda1 (>= $${nvidia:MinVersion}) [!armhf !ppc64el] |
package_libcuda1.Debian += libnvidia-tesla-495-cuda1 (>= $${nvidia:MinVersion}) [amd64 i386 arm64 ppc64el] |

package_libcuda1.Ubuntu += libnvidia-compute-495 (>= $${nvidia:MinVersion}) [amd64 i386] |
package_libcuda1.Ubuntu += libnvidia-compute-495-server (>= $${nvidia:MinVersion}) [amd64 i386] |

package_libnvidia-ml1.Debian += libnvidia-ml1 (>= $${nvidia:MinVersion}) [!armhf !ppc64el] |
package_libnvidia-ml1.Debian += libnvidia-tesla-495-ml1 (>= $${nvidia:MinVersion}) [amd64 i386 arm64 ppc64el] |

package_libnvidia-ml1.Ubuntu = $(package_libcuda1.Ubuntu)
-----snip-----

package_libnvidia-ml1.Ubuntu is missing the libnvidia-ml1 line. It's depending upon a specific version of the package available package build time. The Ubuntu Nvidia driver packages provide libnvidia-ml1 now, but might not have in the past.

[ash ketschup (ashketschup)]

I have the same issue and it is almost one year since this bug was reported. I used ubuntu 20.04 before 22.04 and everything worked fine before.

I tell you never touch a running system xD.

Okay I have to confest, that the installation routine was more complicated in the old ubuntu version but at least it worked.

I have the same issue with nvidia-drivers-515 and nvidia-drivers-525.

[ash ketschup (ashketschup)]

Funny Part is that it seemed to work not too long ago, as this guy here is showing in his video.
https://www.youtube.com/watch?time_continue=44&v=8NBgJLZGnPc&feature=emb_logo

[ash ketschup (ashketschup)]

I was able to find a workaround by doing the installation the "complicated nvidia way". Just in case anybody else needs a workaround and has no idea how to do so, here the steps I did (no warranty that it works for you and will not break up your system (for me this happened this happened 5 times but finally I found out that the following steps works very well for me.))

#DOWNLOAD CUDNN tar.gz (yes, in order to do so, you have to create an nvidia developer account -.-'#).
https://developer.nvidia.com/rdp/cudnn-download

# install compatible driver (somehow the cuda installation removes the previous installation)
sudo apt install nvidia-driver-515
sudo reboot

# INSTALL CUDA11 (CUDA 12 is not compatible currently)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
 wget http://developer.download.nvidia.com/compute/cuda/11.0.2/local_installers/cuda-repo-ubuntu2004-11-0-local_11.0.2-450.51.05-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-11-0-local_11.0.2-450.51.05-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-11-0-local_11.0.2-450.51.05-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-11-0-local_11.0.2-450.51.05-1_amd64.deb
sudo apt-key add /var/cuda-repo-ubuntu2004-11-0-local/7fa2af80.pub
sudo apt-get update
sudo apt-get -y install cuda

# check if cuda 11 is installed successfully (if command could be found your fine :))
nvcc

# add cuda to PATH variable
echo "export PATH=/usr/local/cuda-11.0/bin${PATH:+:${PATH}}" >> ~/.bashrc
sudo reboot

# copy the cudnn libs (go to folder with cudnn.tar.gz download first)
tar -xvf cudnn-linux-x86_64-8.7.0.84_cuda11-archive.tar.xz
sudo cp cudnn-*-archive/include/cudnn*.h /usr/local/cuda/include.
sudo cp -P cudnn-*-archive/lib/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*

# install pip (to) install tensorflow
sudo apt install python3-pip

# install tensorflow 2.9.1 since 2.11 did not work for me with cudnn
pip install tensorflow==2.9.1

# create a symlink to libcusolver.so.11
sudo ln -s /usr/local/cuda/targets/x86_64-linux/lib/libcusolver.so.11 /home/horus/.local/lib/python3.10/site-packages/tensorflow/python/libcusolver.so.11

# check if gpus are available
python3 -c 'import tensorflow as tf; print(tf.version.VERSION);print(tf.config.list_physical_devices());'