HPC and Academy¶
ChemGraph separates agent logic from execution so chemistry tasks can run locally or through a facility backend. HPC integrations require site modules, allocations, scheduler policies, endpoint identifiers, credentials, and shared-filesystem planning beyond a normal PyPI install.
Execution backends¶
| Backend | Extra | Typical role |
|---|---|---|
| Local | Core | Validate task packaging and results on one machine |
| Parsl | parsl |
Submit through a configured Parsl executor |
| Ensemble Launcher | ensemble_launcher |
Launch ensembles on supported systems |
| Globus Compute | globus_compute |
Invoke registered remote endpoints |
| Globus Transfer | Core SDK plus credentials | Move inputs/results between collections |
Start with the execution demos. They separate direct backend calls from agent-driven variants. Test the direct path first so infrastructure failures are isolated from LLM behavior.
Backend selection can use environment variables or [execution] settings. Do
not copy another user's endpoint IDs, allocation names, or private paths.
Academy campaigns¶
Academy supports persistent multi-agent campaigns and a dashboard runtime:
python -m pip install "chemgraph[academy,parsl,globus_compute]"
chemgraph academy --help
chemgraph dashboard --help
The MACE ensemble-screening example has a README and end-to-end guide. Review campaign JSON/JSONC, runtime profiles, prompt profiles, and data paths before launch.
ALCF model access and MCP¶
Compute nodes may not reach the same authentication services as login nodes. Follow the current Argo connection example and ALCF policy for proxies, tokens, outbound networking, and secrets.
ChemGraph also includes ASE, MACE, gRASPA, and XANES HPC MCP modules with direct or Parsl execution. They are deployment building blocks and can contain facility-specific assumptions. See MCP servers.
Operational checklist¶
- Validate the scientific command directly on the target system.
- Validate the backend without an agent.
- Confirm input/output visibility across hosts and containers.
- Run one small agent task with explicit calculator/resource settings.
- Add concurrency only after logs, retries, and cleanup are understood.
- Record model, calculator, environment, scheduler, and code versions.
Never place tokens in campaign files, prompts, repository config, or job logs. Use facility-approved secret mechanisms and least-privilege endpoints.