load_cudart_so fails on systems with both CUDA 12 and CUDA 13

[BlackSamorez]

Since load_cudart_so checks for both libcudart.so.12 and libcudart.so.13 and hard-fails if it sees both from dlopen, cudnn simply wouldn't launch on machines with both CUDA 12 and CUDA 13 install on either system level or environment level.

Is that the intended behavior?

[Anerudhan]

Hi @BlackSamorez ,

Thanks for the question.

Yes, that is by design. Would you help me with a case where both cuda-13 and cuda-12 is installed? Can you not use alternatives in that scenario?

Eg. for ubuntu

sudo update-alternatives --install /usr/local/cuda cuda /usr/local/cuda-12.8 128
sudo update-alternatives --install /usr/local/cuda cuda /usr/local/cuda-13.1 131

sudo update-alternatives --config cuda

In this scenario, only one of the cuda is installed and recognized as expected.

Thanks
Anerudhan