Instructor guide

Prepare the model, ML-IAP export, ALCHEMI SIF, native MPI-LAMMPS runtime, and private results directory before the session. Their source and content identities must be reviewed together; a filename alone is not enough. Keep the actual .env outside Git. Test the exact artifact set in a short GPU allocation, then end that allocation.

For a live class, obtain one GH200 allocation under current site rules. Load and select the private Jupyter environment exactly as in the Arrhenius setup, then load .env and run bash scripts/start-jupyter.sh from the repository root. It starts only JupyterLab, on compute-node loopback, with token authentication. It does not submit a job or start an SSH tunnel. A session-specific connection script should establish the permitted route to your computer; never reuse a previous allocation’s node, TLS certificate, or token.

The Sphinx site can be built without an allocation:

source /path/outside/git/mlip-notebook-venv/bin/activate
make html
make livehtml PORT=8766

make livehtml serves the published pages on the machine’s loopback and rebuilds them when Markdown changes. It is not the runnable JupyterLab server. Forward its port over SSH if viewing it from another computer.

Run the GPU notebooks in order: B4, B5, then B6 (one-rank LAMMPS). Execute only one GPU notebook kernel at a time. B7 works without a GPU and is safe to inspect before class. B2, B3 and B8 (builds and multi-GPU) are optional; give learners prepared artifacts when build or queue time would dominate the session.

Short live demo (webinar)

For a short live demo of Part A, book a one-GPU interactive allocation well before the session; queue wait is the main risk. Install the pixi environment and download the checkpoint on a login node beforehand; compute nodes have no internet access. Set PS1='$ ' so the prompt shows no user, host or path. After the run, show summary.json rather than scrolling the log. If the allocation or run fails, show the last saved summary.json from a rehearsal and state that it is a saved result. Serve the lesson pages locally with make livehtml, which binds to 127.0.0.1, rather than sharing a remote session.

If the site is in maintenance on the day, walk through saved runs instead and say so.

Publication

Before publication, perform a privacy review of Markdown, scripts, notebook outputs, metadata, and rendered HTML. Do not publish usernames, personal hostnames, project/account numbers, tokens, site-private paths, scheduler logs, or model weights. Remove notebook outputs generated on a live site; the source .md files, not .ipynb exports, are the maintained notebooks.