Deployment on Docker
The recommended shape for a server. See Docker quickstart to get running; this page is the production detail.
Image selection
Section titled “Image selection”| Target | Tag |
|---|---|
| NVIDIA | iterateai/lifeboat:latest |
| AMD MI210/MI250 | :latest-rocm / :amd |
| AMD MI300X/MI325X | :amd-mi300x |
| AMD MI350X/MI355X | :amd-mi355x |
| No GPU / integrated / Arm | :lite |
All tags: hub.docker.com/r/iterateai/lifeboat. Version tags such as :2.2.44 are immutable; :latest, :latest-rocm and :lite roll forward.
Pin a version tag in production and upgrade deliberately.
The control plane serves the console, the management API and the inference proxy on one configurable port (8001 by default). Each model server listens on its own port behind that proxy — those do not need to be exposed outside the host, and generally should not be.
Node agents listen on their own port for control-plane commands.
Volumes
Section titled “Volumes”Persist two paths:
- Data directory — registry, users, keys, audit log, licence, logs, download cache. Losing this changes the cluster identity a licence is bound to, which means the licence must be re-issued.
- Models directory — weights. Large.
Mounting the log path to a host volume gives longer retention than the container’s own life.
GPU access
Section titled “GPU access”NVIDIA — the container toolkit, plus a one-time host-setup step that configures device passthrough and installs a refresh service so reboots and driver upgrades do not break GPU access. It is idempotent and safe to re-run.
AMD — the ROCm device nodes are passed into the container directly, with the group memberships that let the unprivileged user inside open them. There is no toolkit and no host-setup step. AMD’s own container guidance requires a relaxed seccomp profile for ROCm’s userspace device calls; that is their requirement, not a Lifeboat choice, and it is the one place the AMD posture is weaker than NVIDIA’s.
No GPU — nothing.
Shared memory
Section titled “Shared memory”LIFEBOAT_SHM_SIZE (default 8 GB) matters only for tensor parallelism above 1, and is fixed at container creation.
In front of it
Section titled “In front of it”Put your own reverse proxy or ingress in front for TLS, hostname routing and access control. Lifeboat can also terminate TLS itself given a certificate and key. Point Prometheus at /metrics and ship logs to your existing platform.
Upgrading
Section titled “Upgrading”Stop, pull, start, keeping the data volume. See Upgrading.
Sources and references
Section titled “Sources and references”- Variables: Environment variables
- Kubernetes instead: Deployment on Kubernetes
- Security posture: Security posture