[Solved] How to build PyTorch on a compute server

(This was originally a question but I found a solution)

Those steps worked for me:

  • If your compute server has a CUDA module that you need to load using module load, DON’T load it before building pytorch (this may not be needed)
  • Set $CUDA_HOME to where cuda is installed. (for me it was: /software-gpu/cuda/8.0.44)
  • Install torchvision BEFORE building pytorch, as with conda torchvision installs cudatoolkit and this screws up the CUDA paths and cause an error when doing import torch
  • Run python setup.py clean before python setup.py install