PyTorch should be installed from Torch repo, not from PyPi #98
Labels
kind/feature
New feature or request
priority/minor
An issue with the product that doesn't impact the user much or not at all (ie tech debt)
Milestone
As PyTorch comes with its own embedded CUDA, we have to make sure we match the CUDA version (and Toolkit) we install in the image with the one from PyTorch. The only way to that is to install from the original PyTorch repo: https://download.pytorch.org/whl/torch_stable.html
We can then specify the exact version we want, like torch==1.9.0+cu111, which would install Torch with CUDA11.1.
The text was updated successfully, but these errors were encountered: