Docking on Modal

Run the GPU docking service (autodock-gpu) on Modal's serverless GPUs, so dock_molecules works without you owning a card. Docking is synchronous and short (~30–60 s per ligand), which fits Modal's request/response + scale-to-zero model well: you pay for GPU-seconds per docking, not for an idle GPU.

Prerequisites

• A Modal account and the CLI: pip install modal then modal setup.
• The engine running somewhere (laptop or cloud). You will point it at the Modal URL.

Deploy the docking service

Modal can run an existing image with Image.from_registry and proxy HTTP to a server running inside it. The autodock-gpu image already starts its FastAPI server on port 8022, so the Modal app just needs to launch that and expose the port. Save this as autodock_modal.py:

import modal

# The published image already contains AutoDock-GPU + its HTTP server.
image = modal.Image.from_registry("ghcr.io/novomcp/autodock-gpu:latest")
app = modal.App("novomcp-autodock", image=image)

@app.function(gpu="A10G", timeout=900, scaledown_window=120)
@modal.web_server(8022, startup_timeout=180)
def serve():
    # Start the image's own server on 8022. Replace the command below with the
    # image's documented entrypoint/CMD if it differs (check the repo's
    # Dockerfile) — Modal proxies whatever listens on the declared port.
    import subprocess
    subprocess.Popen(["python", "-m", "uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8022"])

Deploy it:

modal deploy autodock_modal.py
# -> https://<workspace>--novomcp-autodock-serve.modal.run

Wire into the engine

export AUTODOCK_GPU_URL=https://<workspace>--novomcp-autodock-serve.modal.run

Verify

curl -s "$AUTODOCK_GPU_URL/health"
# {"status":"healthy","service":"autodock-gpu","gpu_available":true}

The first call cold-starts a GPU container (~30–60 s); calls within the scaledown_window are warm. Because docking is a single request/response, the cold start lands on the first dock_molecules and not again until the container scales down.

Cost and access

• You pay per GPU-second. An A10G docking (~30–60 s) is a few cents, and scale-to-zero means no idle cost between jobs. Pick a smaller GPU (T4/L4) to cut cost, or a bigger one (A100) for large libraries.
• Security: a Modal web URL is public, and the docking service has no auth of its own — anyone with the URL can run GPU jobs on your account. Protect it with Modal proxy auth (requires_proxy_auth=True on the web server, plus a proxy-auth token) or keep the URL private, and set a Modal spend limit.

Notes

Verified against Modal's current docs (September 2026): existing images, GPU acceleration. The exact server-start command depends on the image's entrypoint — adapt the subprocess line to match. This guide is reference-quality and has not yet been live-tested end to end.