I’m wanting to deploy TF on a newer system which defaults to Cudatoolkit 13.2. Wanted to see the possibility of this?
I have the same issue with CUDA 13.3 and TensorFlow 2.21.0 on Debian 13.
This compatibility matrix shows that TensorFlow 2.21.0 requires CUDA 12.5:
My expectation was that CUDA 13.3 provides backward compatibility for CUDA 12.5 applications. Therefore, I do not understand why TensorFlow 2.21.0 won’t work with a CUDA 13.3 environment.
Maybe somebody could be so kind and explain the relationship between TensorFlow and CUDA versions.