Server Mode

Long-running HTTP service for repeated energy/force requests. Useful with workflow managers and ORCA external optimization.

Commands

Start server:

enerzyme listen -c server.yaml -m model_dir/ -o server_out/ -b 0.0.0.0:5000 -mc config.yaml

Client request:

enerzyme request -u http://127.0.0.1:5000 -f ORCA -i input.extinp.tmp -k FF02

Shutdown:

enerzyme kill -u http://127.0.0.1:5000

Implementation

• enerzyme/listen.py — Flask/Waitress, route POST /calculate

• enerzyme/tasks/server.py — model forward pass

• enerzyme/request.py — ORCA .extinp.tmp → .engrad bridge

Request format

JSON body (conceptual):

{
    "model_key": "FF02",
    "input_file": "/path/to/geometry",
    "features": {
        "Ra": [[...]],
        "Za": [...],
        "N": 100,
        "Q": -1
    }
}

Exact schema follows what Server.calculate expects for your model’s active features.

Response

JSON with outputs (energy, forces, etc.) and units (Hartree_in_E, Bohr_in_R).

Multi-model serving

listen loads all active: true models from config.yaml. Clients select via model_key (-k).

Server config

No dedicated listen.yaml in the repository. A minimal config may only need:

Datahub:
    preload: true

Model architecture and transforms come from -mc / model_dir/config.yaml.

Deployment notes

• Model load time is paid once at startup — amortize over many requests

• Bind address -b controls network exposure

• Logs go to out_dir and waitress logger (wired to Enerzyme logger)

• For batch evaluation on static datasets, prefer enerzyme predict