I recently set up a FastAI VM on Google Cloud Platform (GCP) using this guide. In the process, I discovered an issue that prevents CUDA from being properly installed. This manifested itself as torch.cuda.is_available()
returning False
in the lesson 1 Jupyter notebook
Due to this Ubuntu bug, running the installer mentioned in the guide will not properly install the Nvidia drivers. The solution is to run this command after the initial reboot:
sudo apt-get -o Dpkg::Options::="--force-overwrite" install --fix-broken
After another reboot, torch.cuda.is_available()
should return True
, and the Jupyter notebook should work as expected.