Hi! I have a system that has CUDA 10.0 (and can’t be changed easily) and want to update our PyTorch version to one of the latest version (v1.10.2). The problem is that when compiling Pytorch, there is a hard requirement for CUDA >= 10.2 at pytorch/cuda.cmake at master · pytorch/pytorch · GitHub
I am wondering if that’s indeed a hard requirement or if it could be compiled with CUDA 10.0 without any problems.
Since a fatal error is raised, I would recommend to keep the check. Note that you are checking the cmake files which would be used in a local source build. If you want to install the pip wheel or conda binaries for 1.10.2, both would ship with their own CUDA runtime and do not use the local CUDA toolkit unless you build custom CUDA extensions.