Torch.cuda.is_available() == True using Python 3.7.1, but False on Python 3.6.9


I’m trying to run some code on a VPS that I don’t have many permissions for (and I won’t be able to get requisite permission). For some reason, I am not allowed to install packages for the Python 3.7.1 installation on this machine, but I can do so just fine for the Python 3.6.9 installation.

I have gotten my repository set up properly with all required packages, and my code runs, but when I try to train my model, it runs on the CPU. Upon further investigation, I found that opening a Python 3.6.9 interpreter and running torch.cuda.is_available() returns False. However, opening a Python 3.7.1 interpreter and running the same command returns True. Furthermore, in the 3.6.9 interpreter, torch.backend.cudnn.enabled returns True. Is there any way I can fix this? I am running torch 1.4.0, nvidia-smi v410.104, and nvcc release 9.1, V9.1.85.


How did you install both PyTorch versions for these different Python versions?
Did you install the initial CUDA version in your Python3.7 environment?