I am installing Pytorch 1.8.1 in a new Conda env on Ubuntu machine with multiple Nvidia 2070 Super and from the Pytorch homepage, we can either install the package cudatoolkit=10.2 or cudatoolkit=11.1.
nvidia-smi shows
Driver Version: 450.102.04 CUDA Version: 11.0
How does one generally decide which version of CUDA to use?