Before you start

Get the complete lesson repository first. It contains the Markdown notebooks, scripts, build definitions, dependency locks, and example inputs:

git clone https://github.com/ENCCS/mlip-hands-on.git
cd mlip-hands-on
cp .env.example .env

Edit only .env with paths for your project and prepared artifacts; it is ignored by Git. The source checkout does not include MACE weights, SIF images, LAMMPS binaries, site credentials, or private run outputs. It does include the two small reviewed-results CSVs used by page B7 (Arrhenius and JUPITER). Keep those outside the checkout and check their identities before a run. The commands in later pages assume your shell starts in the repository root (the directory containing examples/ and scripts/). For a Jupyter session, start the server with that same repository as its file root so its .md pages and included source files are visible.

Part B examples use one pinned MACE checkpoint but separate software environments: JupyterLab for the notebooks, an ALCHEMI SIF on Arrhenius (or a private native environment on JUPITER), and native Python/LAMMPS for ML-IAP/Kokkos. Keeping them separate avoids loading two incompatible CUDA or MPI stacks into one notebook kernel. A1 and A2 use a pixi environment; A3 and A4 use a venv on the LUMI CSC PyTorch module; see their pages.

The core exercises need one GPU and prepared artifacts. Building a SIF and LAMMPS is covered in B2 (ALCHEMI image) and B3 (native LAMMPS), and can be done before the GPU session.

Important

No notebook submits a Slurm job. Obtain an allocation and check the selected model and runtime artifacts before running GPU cells.

Choose a site

The tabs below point to different tested routes; they do not make the sites interchangeable. Selecting a site under Prepare also selects it under Run. The complete variable list is in Environment variables.

Prepare

Read the Arrhenius setup for the ALCHEMI SIF and native MPI-LAMMPS runtime. This is the complete ALCHEMI-and-LAMMPS notebook route.

Read the JUPITER setup for its separate native ALCHEMI, ML-IAP/Kokkos, and Metatomic environments. Do not reuse the Arrhenius SIF or MPI binary.

Read the Leonardo setup for its bounded A100 single-GPU route. It does not qualify the complete LAMMPS notebook sequence.

Run

After checking the artifacts and obtaining one GPU, run the silicon trajectory, ALCHEMI batch, and LAMMPS replicas in order.

Use the JUPITER-specific job and notebook instructions. The ALCHEMI batch notebook has passed a short native functional check; the full LAMMPS notebook route remains Arrhenius-specific.

Use only the tested Leonardo single-GPU commands unless a separate site check qualifies a broader route.

Note

The lesson uses NVIDIA ALCHEMI Toolkit, not the separately packaged ALCHEMI NIM service.