Component

core/ai-model/docker-model-runner

A model served by Docker Model Runner, named by its OCI model reference (`ai/smollm2`, say), behind an OpenAI-compatible API. The same reference is served on both targets, so one Aggregate builds for Docker Compose and for Kubernetes and asks both for the same model. The reference is a deterministic pointer the runner resolves when it pulls the model; the compiler never fetches it. On Docker Compose the host's own Docker Model Runner serves it: the compose file declares the model in its top-level `models` element, and Compose pulls and configures it when the stack starts. That needs what no build can check: Docker Compose 2.38 or later, Docker Model Runner enabled on the host, and a model CLI of at least 1.0.6 for `runtime_flags` to reach the engine. On Kubernetes the Docker Model Runner server runs in-cluster, pulls the same reference into a per-replica model cache, and serves it behind a ClusterIP Service. Consumers read the endpoint, model, host, port and protocol over a `uses` edge. Each operator setting works on the target where it means something and is refused by name on the other.

Provenance

OfficialSigned by a signing identity this catalog recognizes as an official publisher.
Content digest
sha256:5aaacb32d0317067ed01adb05a62cc0513b1a0f64ff8d769d91c431cbfe82680
Published
2026-11-10T10:09:13Z
Signing identity
https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

Signature

  • Officialapplication/vnd.dev.sigstore.bundle.v0.3+json

    https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

    sha256:46e2e3962478eb2bae9e609cc44ff00d2d8b437a57513154a25a42bbc1b47a1f

Reference at this version

oci://hub.aggrega.ai/component/core/ai-model/docker-model-runner:1.0.0

Developer parameters

Values the developer sets in the Aggregate.

Extends BaseComponents#/schemas/BaseModelParameter

NameTypeDefaultRequiredDescription
runtime_flagsarray—noRaw command-line flags passed to the inference engine when the model is started, in Compose's own `runtime_flags` sense. Each flag reaches the runner as one list item and is never read by a shell. On Docker Compose they reach the engine only where the host's model CLI is at least 1.0.6; an older one drops them without saying so.

Operator sizing

Infrastructure sizing the operator tunes per environment.

NameTypeDefaultRequiredDescription
endpointstring—noBase URL of the OpenAI-compatible API of the host's Docker Model Runner, as a container on the compose network reaches it. Read by the Docker Compose realization only, which defaults it to Docker Desktop's container-facing address, http://model-runner.docker.internal/engines/v1; on Docker Engine set it to the address the host's runner answers containers at, such as http://172.17.0.1:12434/engines/v1. The build pipes the address and never resolves it. The Kubernetes realization publishes its own Service address and refuses this key.
portinteger (min 1024, max 65535)—noPort the in-cluster runner listens on and its Service accepts requests on. Read by the Kubernetes realization, which defaults it to 12434; on Docker Compose the port is the host runner's own and part of infra.endpoint, so the key is refused there.
engine_versionstring—noDocker Model Runner image tag the Kubernetes realization runs (default v1.2.8), on the server and on its model-pull step; with a GPU requested the CUDA variant of that tag is run. On Docker Compose the runner is the host's own installation, so the key is refused there.
replicasinteger (min 1)—noNumber of serving replicas on Kubernetes (default 1), each with its own model cache and its own pull. On Docker Compose one host runner serves the model, so the key is refused there.
computeobject—noCPU and memory for the serving container on Kubernetes (defaults 1 and 4Gi), as requests and limits. On Docker Compose the host runner is sized by the host, so the key is refused there.
gpuobject—noOptional GPU scheduling for the serving container on Kubernetes. When count is set, the Recipe requests GPUs as the extended resource nvidia.com/gpu and runs the runner's CUDA image; a named runtime class schedules the pod under it. Omit the whole object to serve on CPU. On Docker Compose the host runner decides its own GPU use, so the key is refused there.
model_cacheobject—noPer-replica model store on Kubernetes (default size 20Gi), sized for the pulled model; each serving replica gets its own claim. On Docker Compose the host runner keeps its own store, so the key is refused there.

Output contract

Values this component publishes to consumers over the uses topology.

NameTypeDefaultRequiredDescription
hoststring—yesHost name the runner's API is reached at: the in-cluster Service name on Kubernetes, and the host part of the endpoint on Docker Compose. A consumer that composes its own address dials this name; one that does not reads `endpoint` instead.
portinteger—yesTCP port the runner's API accepts requests on: the Service port on Kubernetes, and the port of the endpoint on Docker Compose (its scheme's default where the endpoint names none).
endpointstring—yesFull base URL of the OpenAI-compatible API the runner serves, with the host, the port and the API path already composed. A consumer hands it to a client library as its base URL.
modelstring—yesName a request must ask the runner for, which is the OCI model reference as the Aggregate wrote it, on both targets.
protocolstring—noURL scheme to reach `host` and `port` over, for a consumer composing its own address rather than reading `endpoint`.

Uses sockets

The dependency sockets this component connects through. The target of every connection wired through a socket must satisfy its label selector, and each connection may carry an attachment in the slots the socket declares. A required slot is owed by each connection the Aggregate or the Release wires through the socket.

No uses sockets declared.

Attachment slots

Named binding points at which a Feature attaches to this component. A slot's selector constrains which Feature may fill it and its inject list names how the Feature's material may be delivered. A required slot must be bound on every instance, by an attachment on the instance or at a scope above it, and a sensitive slot is declared to handle secret material. A Feature may also attach under a name no slot declares.

No attachment slots declared.

Self-inject allowlist

env

Relationships

Typed edges to and from other published entities. A reference with no published target renders as a broken edge rather than a link.

Uses

Used by

Component concept reference →

The Aggrega catalog — an open index of components, features, recipes, and the plugins that render them.

Built with the shared Aggrega design system.

Privacy