This issue seems to indicate that the CUDA path can be specified by running say
CUDA_HOME=/path/to/cuda python3 setup.py install. I have CUDA 8 available in /path/to/cuda and an ancient version in /usr/local/cuda. When I run the above command, PyTorch says that it has found the ancient version in /usr/local/cuda and that it, of course, is too old.
What is the correct way to specifiy the only CUDA path that should be considered?