Component

core/ai-model/vllm

vLLM running in-cluster as a self-hosted inference server. A Recipe synthesizes the runtime directly: a weights-pull initContainer resolves the model pointer into a model-cache volume the serving container mounts, and a ClusterIP Service fronts the OpenAI-compatible API. The model value is a deterministic pointer resolved at pod start, never fetched at compile time. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint. Each realization also publishes where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, where one address cannot reach every replica, the Service or the pod labels a scraper discovers them by. On KServe the cluster's serving runtime decides the metrics port and path, so nothing is published there. On a cluster running the Prometheus Operator, `core/prometheus-exporter` attached to the synthesized Kubernetes realization, its port set to `http`, emits a ServiceMonitor for it instead.

Provenance

OfficialSigned by a signing identity this catalog recognizes as an official publisher.
Content digest
sha256:b13caeae34054f971912cfb864d53c9efbe299e59dbd5280ed7ce312e50fed52
Published
2026-11-10T10:09:41Z
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

    sha256:50a6d28f11d3c84b4ec7b3b89bbdecc5d9f1523323fda172c2340fbd633553ce

Reference at this version

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

Developer parameters

Values the developer sets in the Aggregate.

Extends BaseComponents#/schemas/BaseModelParameter

NameTypeDefaultRequiredDescription
served_model_namestring—no—

Operator sizing

Infrastructure sizing the operator tunes per environment.

NameTypeDefaultRequiredDescription
portinteger (min 1024, max 65535)8000noServer listen port for the model API. Most operators should omit this and let the Recipe assign the default (8000).
engine_versionstringv0.6.3novLLM runtime image tag. Selects the image tag on the synthesized serving and weights-pull containers.
replicasinteger (min 1)1noNumber of serving replicas. Each replica holds its own model-cache volume and pulls the weights independently.
min_replicasinteger (min 1)—noLower bound for the replica count when horizontal autoscaling is configured. Declared ahead of rendering — the synth recipes serve a fixed replica count and do not yet emit an autoscaler; set it to record the intended lower bound.
max_replicasinteger (min 1)—noUpper bound for the replica count when horizontal autoscaling is configured. Declared ahead of rendering — the synth recipes serve a fixed replica count and do not yet emit an autoscaler; set it to record the intended upper bound.
computeobject—no—
model_cacheobject—noPer-replica model-weights cache volume. Sized for the pulled model; each serving replica gets its own claim.
model_artifactstring—noOptional OCI reference to the model weights, an alternative weights source to the Hugging Face model id in values.model. A deterministic pointer the runtime resolves at pod start; the compiler never fetches it. Declared ahead of rendering — the synth recipes resolve weights from values.model and do not yet consume this pointer.
control_plane_presentbooleanfalsenoOperator affirmation that the model-serving control plane (the KServe or KubeAI controller and its custom-resource definitions) is installed in the target cluster. A controller-driven realization halts the build unless this is set true, so a serving custom resource is never emitted as an orphan into a cluster that cannot reconcile it. The direct synthesis recipes ignore this flag.
overridesobject (open)—noProvider-specific overrides. Accepts any nested structure. Each Recipe documents which keys it reads.
gpuobject—noOptional GPU scheduling for the serving container. When set, the Recipe requests GPUs as the extended resource nvidia.com/gpu and, when a runtime class is named, schedules the pod under it. Requires a GPU-capable cluster and, for runtime_class, a matching RuntimeClass configured in the cluster. Omit the whole object to serve on CPU.

Output contract

Values this component publishes to consumers over the uses topology.

NameTypeDefaultRequiredDescription
hoststring—yesIn-cluster DNS name of the Service fronting the inference server. Which Service that is follows the realization the Manifest selected, so a consumer reads this rather than deriving a name.
portinteger—yesTCP port that Service accepts model requests on. A controller-driven realization fronts the server on its own port, so a consumer reads this rather than assuming a default.
endpointstring—yesFull base URL of the OpenAI-compatible API the server 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 this server for. It is the served name the realization registered, which is not always the weights id the Aggregate declared.
protocolstring—noURL scheme to reach `host` and `port` over, for a consumer composing its own address rather than reading `endpoint`.
metrics_endpointstring—noThe full address Prometheus scrapes this server's metrics at — scheme, host, port and path, `http://<host>:<port>/metrics`, on the port the API is served on. Empty where one address cannot reach every replica, which then publishes the service or pod facts instead, and on KServe, which publishes none.
metrics_servicestring—noName of the Service every replica's metrics sit behind, for a scraper that discovers each endpoint of that Service rather than scraping one address. Set by the Kubernetes synthesis while infra.replicas is above one, beside `metrics_service_port` and `metrics_path`; empty otherwise.
metrics_service_portstring—noName of the port of `metrics_service` that carries the metrics, `http`. Set and empty together with `metrics_service`.
metrics_pod_selectorstring—noKubernetes label selector matching every pod that serves this server's metrics, for a scraper that discovers those pods directly because no Service of the realization carries them. Set by the KubeAI realization, beside `metrics_pod_port` and `metrics_path`; empty otherwise.
metrics_pod_portstring—noName of the container port on those pods that carries the metrics, `http`. Set and empty together with `metrics_pod_selector`.
metrics_pathstring—noPath the metrics are served at, `/metrics`, published beside the service or the pod facts. Empty where the server publishes `metrics_endpoint`, which carries its own path.

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.

SlotSelectorRequiredInjectSensitiveDescription
model_pull_tokenfeature.aggrega.ai/kind=secretnoenvyesOptional model-registry pull credential for gated model weights. When attached, the weights-pull initContainer receives it as an environment variable; unattached, weights are pulled anonymously.
api_keyfeature.aggrega.ai/kind=secretnoenvyesOptional key the server requires of every caller. When attached, the Recipe delivers it as VLLM_API_KEY, the variable vLLM reads at startup, and each request under /v1 must then carry it as a bearer token in the Authorization header. The server checks this one key, so every caller must present the same key. Unattached, the server requires none.

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