I’m on Debian 12 trying to get a certain app working. It uses a tensorflow version that requires cuda 11. I already have cuda 12.5 installed, so I downloaded the cuda 11.8 run file installer, installed it to /usr/local/cuda-11.8, and changed my ~/.bashrc from using /usr/local/cuda to /usr/local/cuda-11.8. The app with the tensorflow is still saying it can’t find libcudart.so.11.0 and other libraries

This is my setting on ~/.bashrc. I’ve also rebooted just to be sure, but it’s still not working.

export CUDA_HOME=/usr/local/cuda-11.8
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
export PATH=$CUDA_HOME/bin:$PATH

nvcc –version is already displaying 11.8. I checked that the libraries do exist in /usr/local/cuda-11.8/lib folder. The system-wide /etc/ld.so.conf.d/cuda-11-8.conf also exists and points out to the correct path.

sudo find / -name 'libcudart.so.11.0'
/usr/local/cuda-11.8/targets/x86_64-linux/lib/libcudart.so.11.0

The only notable changes I did during cuda 11.8 installation is not installing the demo kit and documentation files, not installing the symlink to /usr/local/cuda because I already have it pointing to cuda 12, and not installing the driver because I already have the correct one installed.

Do I really have to wipe cuda 12 to get cuda 11.8 working just for this one app?