Hold up. So you are saying it doesn’t even matter what the default CUDA version (nvcc --version) on my system is? I always thought that when you install PyTorch (from the “get started” page), you had to select the Cuda version that is currently installed/the default on your system. But from what you are saying, it seems that that doesn’t even matter and that pytorch installs the correct CUDA libs in the environment? I never realized this! What does that mean for other scripts in the environment though? Will they all use the PyTorch loaded CUDA libs (so does PyTorch set some env specific variables to CUDA so the current env knows where those libs are?).