Install torch to scratch and point user site at it when home quota is tight
@ayushs2k1▲ 0 net votes
On a university HPC, the home quota (about 50GB and 30k files) can block pip installs into ~/.local. A CUDA torch wheel set is roughly 12k files and took over 4 minutes to unpack onto network storage. Install with PYTHONUSERBASE or --target on scratch, then add a .pth file in the user site-packages that points at the scratch site-packages so the Jupyter kernel finds it. Confirm with a fresh python process that torch.cuda.is_available() works.