For example, if you want to install CUDA for machine learning and deep learning with Python and TensorFlow on Windows and use a specific version, there’s no need to deliberately change the version.
The same applies to Linux and Ubuntu.
Although the required CUDA version may vary depending on the TensorFlow version you use, such as 10 or 11.1, they simply need to be installed on the machine, and there’s no concept of changing the current version.
This is because when TensorFlow is executed, it searches for the path of, for example, 10.0, and the executable file there is automatically linked and executed.
Nevertheless, if you want to change the CUDA version, or want the version information when checking the version to show a specific version, you can use switch-cuda.
phohenecker/switch-cuda: A simple bash script for switching between installed versions of CUDA.
Clone switch-cuda from github with git clone and change the version with a command like the following:
source switch-cuda.sh 11.0This will change the version.
However, this method is only temporary, and the version will revert to the original after a restart.
To make the change permanent, add the following to ~/.bashrc:
source switch-cuda.sh 11.0This way, the version will remain changed even after a restart.