How to change CUDA version

Question

I met this error when compiling a modified caffe version.

OpenCV static library was compiled with CUDA 7.5 support. Please, use the same version or rebuild OpenCV with CUDA 8.0

I have some old code may not compatible with CUDA8.0, so I want to change my cuda version for this error.

I modified my ~/.bash_profile like this

# export PYTHONPATH=$PYTHONPATH:/usr/local/cuda-8.0/lib64/
# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64
export PYTHONPATH=$PYTHONPATH:/usr/local/cuda-7.5/targets/x86_64-linux/lib/
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-7.5/targets/x86_64-linux/lib/

But it did't work. Still the same error. What should I do?

you need to rebuild opencv with cuda 8.0 there is no other way! — Balaji R, Aug 03 '17 at 08:10
@BalajiR Really? Can't I just change the cuda version? It seems possible according to the error information. — baNv, Aug 03 '17 at 08:19
@BNT I'm not sure which file do you want... Do you mean CMakeLists.txt? Here is the link of the caffe version I am working on. https://github.com/varunjampani/video_prop_networks — baNv, Aug 03 '17 at 08:26
I had a same problem. My solution is changing the bash file like above and remove the CMakeCache.txt — irt, Mar 30 '18 at 02:02

score 30 · Answer 1 · answered Nov 20 '21 at 23:41

30

Perhaps cleaner:

sudo update-alternatives --display cuda

sudo update-alternatives --config cuda

answered Nov 20 '21 at 23:41

w.t

411
4
5

This is the best answer. Simple commands, no need to install anything else, & reversible. – Jun Dec 02 '21 at 13:12
This is a safer and more standard procedure. Less manual intervention at the command line leads to fewer errors in running specific versions of programs. Just use it consistently to find out what is selected and in what mode (auto goes to the option with the highest priority) and to change it manually if you need to. That way the version of cuda will change at the system level without setting symlinks by hand. – Pablo Adames Apr 11 '23 at 16:42

score 26 · Accepted Answer · edited Sep 14 '21 at 17:26

26

Change your CUDA soft link to point on your desired CUDA version. For example:

ll /usr/local/cuda lrwxrwxrwx 1 root root 19 Sep 06 2017 /usr/local/cuda -> /usr/local/cuda-8.0/

Simply relink it with

Update: If the symlink already exists, use this other command:

[jalal@goku ~]$ ls /usr/local/cuda
lrwxrwxrwx. 1 root root 20 Sep 14 08:03 /usr/local/cuda -> /usr/local/cuda-10.2

[jalal@goku ~]$ sudo ln -sfT /usr/local/cuda/cuda-11.1/ /usr/local/cuda
[jalal@goku ~]$ ls /usr/local/cuda
lrwxrwxrwx. 1 root root 26 Sep 14 13:25 /usr/local/cuda -> /usr/local/cuda/cuda-11.1/

ln -s /usr/local/cuda-7.5 /usr/local/cuda

(With the proper installation location)

edited Sep 14 '21 at 17:26

Mona Jalal

34,860
64
239
408

answered Sep 06 '17 at 07:50

rkellerm

5,362
8
58
95

5

If you use the `ln -s` command, it will fail if the symlink already exists. So remove the old symlink with a simple `sudo rm /usr/local/cuda`. (As always, be very careful with your paths when using `sudo rm`!) – Olivia Stork May 23 '18 at 15:40
1

I don't have the sudo permission, and got the following error ln: failed to create symbolic link ‘/usr/local/cuda-10.0/cuda-8.0’: Permission denied – Shuai.Z Apr 08 '19 at 05:19
7

You can also use `sudo ln -sf` to create OR update the symlink if it already exists. Then you won't need the `rm` command. – Olivia Stork May 14 '19 at 17:31
This doesn't work for me - after a reboot, the symlink is correct, but `nvidia-smi` gives me the wrong version. – chris Jul 31 '20 at 20:47
1

@ChrisAnderson `nvidia-smi` is showing the cuda version for driver API. Use `nvcc --version` to find out cuda version. See here https://stackoverflow.com/questions/53422407/different-cuda-versions-shown-by-nvcc-and-nvidia-smi – yuqli Aug 13 '20 at 19:18
Oddly `sudo ln -sf` doesn't work for me (on an Ubuntu 20.04 based distro, ln version 8.30), I agree it should do... Had no effect. Cannot reproduce in my own namespace without sudo... – Louis Maddox Apr 19 '21 at 15:06
1

@LouisMaddox check the edits I made to the post – Mona Jalal Sep 14 '21 at 17:28
Please remember to update `PATH` and `LD_LIBRARY_PATH` variables after doing this as explained [here](https://stackoverflow.com/questions/63737492/opencv-claims-to-find-wrong-cuda-version/75080353#75080353). – Eljas Hyyrynen Jan 11 '23 at 08:51
1

This should be avoided in favour of using the safer update-alternatives package as the answer by [@w-t](https://stackoverflow.com/users/7223954/w-t) indicates. – Pablo Adames Apr 11 '23 at 16:38

score 13 · Answer 3 · answered Dec 19 '18 at 08:04

13

Maybe a bit late, but I thought it might still be helpful for anyone who comes across this question. I wrote a simple bash script for switching to a different version of CUDA within the current bash session: https://github.com/phohenecker/switch-cuda

answered Dec 19 '18 at 08:04

paho

1,162
9
13

1

Great thank! Your toolkit is really useful. For the most of time, I have no access to the sudo authority. – Fang WU Sep 06 '21 at 09:07
best answer imo. It doesn't even need sudo authority to perform – JonnyJack Oct 05 '22 at 04:06

score 7 · Answer 4 · answered Mar 17 '22 at 14:36

This solution explains how you can have multiple different cuda versions installed, i.e. 10.2, 11.3 and 11.6 and switch between them. It's an extension of @w.t and makes use of update-alternatives.

Afaik, after cuda 11.x the installations on Ubuntu 20.04 cuda installations will be added to the update-alternatives maintenance automatically.

Let's say you installed cuda 10.2, cuda 11.3 and cuda 11.6 (following the official nvidia installation guide: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html). They will all reside in:

/usr/local/cuda-10.2/...
/usr/local/cuda-11.3/...
/usr/local/cuda-11.6/...

Your update-alternatives will have two entries:

$ sudo update-alternatives --query cuda
...
/usr/local/cuda-11-3 - priority 113
/usr/local/cuda-11-6 - priority 116

Solution 1: If you want to make use of the update-alternatives make sure that your cuda symbolic link points to /etc/alternatives/cuda.

# Change the symbolic link target.
$ sudo ln -sfT /etc/alternatives/cuda /usr/local/cuda

# Check the path.
$ ll /usr/local/cuda

lrwxrwrwrwx 1 root root /usr/local/cuda -> /etc/alternatives/cuda/

Now, all that is left is to make sure /etc/alternatives/cuda points to the version you want to use, e.g. 11.3.

You can make that update with:

$ sudo update-alternatives --config cuda

and follow the instructions to change the version.

Check the path:

$ ll /etc/alternatives/cuda

lrwrwrwrwx root root /etc/alternatives -> /usr/local/cuda-11.3

almost done.

And always make sure to load the correct library PATHs in your ~/.bashrc.

Solution 2: Directly set your /usr/local/cuda symbolic link to the correct version.

$ ln -sfT /usr/local/cuda-11.3 /usr/local/cuda

Reboot your machine and double check everything is set properly:

$ nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Mon_May3 19:15:14_PDT_2021
Cuda compilation tools, release 11.3 V11.3.109
Build cuda 11.3.r11.3/compiler.29920130_0

score 1 · Answer 5 · answered Sep 06 '17 at 09:09

I solved the problem finally.

Modifying ~/.bash_profile to change the path to CUDA is the correct way. But when you changed the file, you need to relaunch the bash.

Simply source ~/.bash_profile won't work. Because source will only append the content in the file to the already existed path rather than cover it.

How to change CUDA version

5 Answers5

Linked