In some situations, my colleagues need to run some old object detection code that only supports PyTorch 0.4.0. However, we only have RTX 2080 Ti, which won’t properly run if CUDA 8 is used. Can I compile PyTorch 0.4.0 from source code using CUDA 10?
No worries, I’ll trade you my 1080 Ti against your 2080 Ti for a small fee. 
More seriously, packages for PyTorch 0.4.2 with CUDA 9.2 can be found on the site, but CUDA 10 seems to be first supported by PyTorch 1.0. So you may have some luck compiling PyTorch 0.4.2 with CUDA 10 or you might look at grabbing fixes from the git repo whenever you hit a compile error (I don’t think there should be too many).