Skip to content
Docs

Docker Quickstart

From nothing to a served model on a Docker host.

Prerequisites

Section titled “Prerequisites”
  • Docker Engine 25.0+ (and Compose 2.21+ if you use Compose).
  • NVIDIA hosts: the NVIDIA Container Toolkit 1.17+, and driver 580.65.06 or newer. The image ships the CUDA 13 runtime; an older driver passes every other check and then fails when a model loads.
  • AMD hosts: a ROCm-capable kernel driver. No container toolkit, no host-setup step.
  • No GPU: nothing beyond Docker and about 3 GB of disk.

The one-line installer

Section titled “The one-line installer”
Terminal window
curl -sSL https://raw.githubusercontent.com/IterateAI/lifeboat-releases/main/install/get-lifeboat.sh | bash

It checks the host, detects the accelerator, selects the matching image, writes an environment file and brings the stack up. On a machine with no GPU it selects the Lite image rather than pulling a CUDA image that cannot run.

By hand

Section titled “By hand”

Fetch the Compose file and the example environment file from the releases repository, then set at minimum:

Variable Why
LIFEBOAT_ADMIN_PASSWORD Optional — if unset, the container generates one per install. There is no shipped default.
LIFEBOAT_MODELS_DIR Where weights live. Persist it; models are large.
LIFEBOAT_DATA_DIR Registry, audit log, licence, logs. Persist it — the licence grace clock and cluster identity live here.

Then docker compose up -d.

First boot

Section titled “First boot”
  1. Open the console on the published port (8001 by default).
  2. Create the first administrator — the console asks on a fresh deployment.
  3. The 24-hour grace period starts at first boot; you can serve immediately without a licence. See Licensing.
  4. Register a model and create a server: Your first model and server.

On NVIDIA a one-time host-setup step runs and exits; it is idempotent and safe to re-run after a reboot or driver upgrade.

Verify

Section titled “Verify”
Terminal window
curl -s localhost:8001/api/version           # liveness and version
curl -s localhost:8001/metrics | head        # Prometheus metrics

The console’s Dashboard should list your GPUs with memory and utilisation. If it does not, the container cannot see the devices — check the toolkit on NVIDIA, or that the ROCm device nodes are passed through on AMD.

Ports and volumes

Section titled “Ports and volumes”

The control plane serves the console, the management API and the inference proxy on one configurable port. Each model server listens on its own port behind that proxy. Persist the data directory and the models directory; mounting the log path to a host volume gives retention beyond the container’s life.

Upgrading

Section titled “Upgrading”

Stop the stack, pull, start again, keeping the data volume — registry, users, keys, audit log and downloaded models all carry across. See Upgrading.

Sources and references

Section titled “Sources and references”