As Rosetta just uses libtorch and not the Python wrapping, we've changed to just using extras=torch, and the relevant documentation page is Building Rosetta with TensorFlow and Torch.
For earlier versions of Rosetta (e.g. Rosetta 3.14), you may still need to use extras=pytorch. Aside from that substitution, the documentation link above should still apply.