“My NVIDIA CUDA version is 11.3, and my NVIDIA driver version is 465.19.01. However, I have installed PyTorch 1.13.1 and CUDA 16.0, and surprisingly, they seem to be working together. I’m quite curious about this. Could someone provide an explanation for this unexpected compatibility?”
That is weird, does anyone know it ?
When you have a time machine that gives you access to Cuda 16.0, everything is possible, my friend.
You would probably be better off buying lottery tickets or something…
CUDA 16.0 is not a thing as already mentioned.
You should however note, that the PyTorch binaries ship with their own CUDA dependencies and you would thus only need to properly install an NVIDIA driver.
Your locally installed CUDA toolkit will be used if you build PyTorch from source or a custom CUDA extension.
But if my driver can only support CUDA 11.3, why it also works for 11.6?
CUDA11 introduced minor version compatibility as explained here.