NVIDIA ALCHEMI Toolkit tutorials

Build GPU-accelerated atomistic simulation workflows with NVIDIA ALCHEMI Toolkit.

Start with Module 00: Toolkit quick tour

📦 Toolkit docs · 📄 Toolkit examples

Why Core?

These tutorials teach the core concepts and workflows of NVIDIA ALCHEMI Toolkit. Future additions may cover more topics, deeper lessons, and simulation workflows inspired by real research and development use cases.

Tutorials

Module	What you will do	Time
00 · Toolkit quick tour	Choose research systems and run a complete batched relaxation.	5–10 min
01 · Toolkit fundamentals and data structures	Create AtomicData, pack several structures into a Batch, and work with Zarr data.	30 min
02 · Simulate and compose	Compose a model and control molecular dynamics with a bias hook; optionally debug a failed update.	40 min
03 · Scale a study	Build fixed-stage, FusedStage, and inflight routes for one molecular study.	15–20 min

Prerequisites

Background

• basic Python and NumPy/PyTorch experience
• some familiarity with basic computational chemistry

Hardware

• a CUDA-capable NVIDIA GPU

Running the tutorials

Launch on NVIDIA Brev

Launch the ALCHEMI tutorials on NVIDIA Brev

Run the tutorials locally

Install uv, clone this repository, and run these commands from the repository root.

Create .venv from the exact packages recorded in uv.lock:

uv sync --locked --managed-python --python 3.12.13 \
  --no-default-groups --group kernel --group lab

This installs the tutorial runtime, Jupyter kernel support, and JupyterLab. It leaves contributor test tools out of the learner environment. The managed Python build includes the C headers used when Torch first compiles CUDA code.

Download and verify the model files used by the tutorials:

uv run --locked --no-sync --env-file environment/runtime.env \
  python environment/prewarm_assets.py

Start JupyterLab:

uv run --locked --no-sync --env-file environment/runtime.env jupyter lab

Open the URL printed by JupyterLab and begin with Module 00.

For VS Code, open a notebook and select .venv/bin/python under Select Kernel > Python Environments.

pyproject.toml declares the packages, and uv.lock records their exact resolved versions. environment/runtime-pins.toml records the Toolkit commits and model checksums.

Contributors

• Ryan Reese created the ALCHEMI Toolkit melting-point study.
• Anoushka Bhutani developed companion tutorial challenges.

License

NVIDIA-authored source and course diagrams are licensed under Apache 2.0. The NVIDIA course banner and NVIDIA marks are excluded from that grant.

Third-party packages, model checkpoints, CUDA components, data, and viewer code use their own terms. NCI Atlas data are distributed under CC BY 4.0 with attribution and change notices. See THIRD_PARTY_NOTICES.md and SOURCES_AND_LICENSES.md.