Use the following instructions to launch a short training job on QM9/H. See default.py for the longer configurations that reproduce the results in the paper. It should train at around 21.9 steps/s ...