If I use a low version PyTorch with a high version CUDA, will I meet any troubles?
Yes, you could run into issues, if you are trying to compile legacy PyTorch versions with more recent CUDA versions, since CUDA methods might have been removed, changed etc.
What’s the reason you would need an older PyTorch version?
Why I need an older PyTorch version? Reimplementing deep learning methods proposed a few years ago and the codes are not maintained.