31

I'm working with Ubuntu 16 and Cuda toolkit 10.2.

I'm using installer file: cuda_10.2.89_440.33.01_linux.run.

I understand I need to remove something I previously installed but don't know what exactly.

Any suggestions will be appreciated.

K7AAY
  • 16,864
  • 9
  • 45
  • 77
Albino Cordeiro
  • 491
  • 1
  • 4
  • 7
  • 2
    Why do you expect to need to remove a previously installed package? What was that package? Which Ubuntu is installed; 16.04 or 16.10 ? Are you using virtualization, and if so, which package? Please click [edit] and add that vital information to your question so all the facts we need are in the question. Please do not use Add Comment. – K7AAY Feb 20 '20 at 19:12
  • 2
    I just got the same message. Just in case anyone wonders, right after "driver found," the messages says something like "YOU ARE STRONGLY ENCOURAGED TO ABORT." So that explains the 'need to remove a previously installed package...' – shmim Nov 21 '20 at 00:27

4 Answers4

19

If you only want to install the cuda toolkit to /usr/local, please run:

sudo sh ./cuda_10.2.<spec>.run --toolkit --silent --override

If you need nvidia driver as well, you could try the answer provided by Albino Cordeiro

Running dpkg -l | grep Nvidia and then sudo apt purge ... did the trick.

Warning: above will uninstall your current nvidia driver.

user3026714
  • 191
  • 1
  • 2
10

Just run into this exact problem in Ubuntu18.04. And after some search, my solution is the following:

  1. If you are sure that your driver version matches the cuda-toolkit version you are gonna install, refer to @domainxz answer, add --toolkit to only install toolkit and not to care about nvidia driver. Check this link for compatibility. (Scroll down to CUDA Driver table).

  2. If you want CUDA to handle the compatibility problem for you, you need to uninstall your current drivers. Follow the instructions in the official documentation. After you entered the text screen after reboot, uninstall your previous Nvidia driver and run the cuda runfile. For Ubuntu 18.04, what I did and worked:

    sudo apt-get remove --purge nvidia\*
    sudo apt-get autoremove
    sudo sh cuda-*.run --silient --override

The official documentation is actually very detailed... Just make sure you follow each step and you should be fine.

Community
  • 1
Yuan Tao
  • 201
  • 2
  • 3
8

I managed to solve my problem. I manage a large number of identical servers and usually use a .run file (cuda_10.2.89_440.33.01_linux.run) to install the Cuda toolkit. On this one specific server, I did some experiments a while ago and forgot that I installed the previous version of Cuda differently there (using sudo apt install). Running dpkg -l | grep Nvidia and then sudo apt purge ... did the trick.

Albino Cordeiro
  • 491
  • 1
  • 4
  • 7
0

got in rootm with

$ sudo su

then

# for FILE in $(dpkg-divert --list | grep nvidia-340 | awk '{print $3}'); do dpkg-divert --remove $FILE; done

then

# apt --fix-broken install
spacycookie
  • 1