PyTorch standard Build Configuration (CI version)

I am trying to build pytorch from source exactly as is done for the official releases.
I found this repo:
Apparently, it contains all the scripts needed for CI, building and testing, however it’s missing an entry point for newbies like me.

Which script is the right one to run? Which arguments? Do I need to set up one of the docker containers first?
Please help with this.

EDIT: When trying to follow the instructions found here (, I realized that the prebuilt containers by soumith do not have a version for cuda 10.2 and cuda 11.0.
Which Dockerfile was used to create those, so that I can build an updated version?

EDIT 2: I haven’t verified, but now my guess is that the prebuilt docker containers from soumith are built from the Dockerfile at
It remains unclear to me though, why the README tells to run them with “bash” once, because the seem to be pulled by the script anyway