With Docker

One command starts it from the release image; it is ready when /health answers.

docker run -d --gpus all --shm-size 8g --restart unless-stopped \
  -p 127.0.0.1:8000:8000 -v decisio-data:/data \
  ghcr.io/aminry/decisio@sha256:5f758e5cfbf42892f42f62e037f26376154c3e77c5caf6afdd6ccc1a5cd9c4ee
# answers once the first start has fetched the weights
curl -s http://127.0.0.1:8000/health

Shown, not run: it needs a GPU host with the NVIDIA Container Toolkit, which this check does not have. Its comments are shown above their commands.

Figure · In a container

The image holds the server; the weights live in a volume.

Docker restarts the container when the server exits; the weights stay in the volume.

1. What you need

• A Linux host with one NVIDIA card, sized as on a cloud GPU.
• Docker with the NVIDIA Container Toolkit, so the container can see the card.

2. Run it with Compose

decisio’s compose file, from the release image:

name: decisio

services:
  decisio:
    image: ghcr.io/aminry/decisio@sha256:5f758e5cfbf42892f42f62e037f26376154c3e77c5caf6afdd6ccc1a5cd9c4ee
    environment:
      DECISIO_BASE: gemma-4-31b       # the default (measured on 96 GB, runs on 80 GB); qwen3.6-35b-a3b or gemma-4-12b for the others
    ports:
      - "127.0.0.1:8000:8000"
    volumes:
      - decisio-data:/data
    shm_size: "8gb"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped           # also after a dead engine: the server exits with code 70

volumes:
  decisio-data:

Shown, not run: decisio's compose.yaml, with the release image in place of its local build; not run at this release (an image built from decisio's main came up with the file on a card on 2026-10-01); run record.

Save it as compose.yaml and run docker compose up -d. It is ready when /health answers 200, as a healthy server’s answer shows.

3. Choose what it serves

Setting	What it does
DECISIO_BASE	the model: gemma-4-31b by default, or gemma-4-12b, qwen3.6-35b-a3b, or a decisio model repository
DECISIO_MODEL	a checkpoint instead; it brings the model its config.json declares
DECISIO_HOST, DECISIO_PORT	where the server listens inside the container

Arguments after the image name go to the server, so every server flag works.

4. What to know

• The first start downloads the weights into the decisio-data volume at /data. They are never part of the image, and a restart does not download them again.
• The container runs as a non-root user, and the port is published on 127.0.0.1 only.
• It checks for an NVIDIA GPU first, and exits with an explanation when it cannot see one.
• When its engine dies, the server exits with code 70. Docker restarts a container when its process exits, not when its health check fails, so restart: unless-stopped brings it back.
• After upgrading the image, certify your tasks again (Upgrade).

5. Build the image yourself

To build from source instead, from a clone of the release:

git clone --branch v0.18.1 https://github.com/aminry/decisio
cd decisio
uv build --wheel
docker compose up --build

Shown, not run: it needs a GPU host with the NVIDIA Container Toolkit, which this check does not have.

This builds the image from decisio’s own Dockerfile and compose.yaml, with no image digest to set.

What next

• On a cloud GPU: the same server without a container, under systemd or Kubernetes.
• Requirements and limits: every machine, the context length and what was measured on what.
• Operate: health, monitoring and what to watch in production.