Questa pagina è stata tradotta automaticamente. Leggi l’originale in inglese.

Component

core/ai-model/nvidia-nim

NVIDIA Inference Microservice (NIM) as a self-hosted, GPU-native inference server serving the OpenAI-compatible API. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: every realization runs a NIM image the reader must be entitled to pull, named by the Release in `infra.image`. The catalog names none, and nothing renders until it is set. On Kubernetes three Recipes realize it. `helm/nvidia-nim` assembles the workload itself and fetches nothing at build time: reach for it where no NIM Operator runs. `helm/nvidia-nim-chart` renders NVIDIA's `nim-llm` chart: reach for it where that chart is wanted and NVIDIA's chart repository is reachable from the build. `helm/nvidia-nim-service` emits a `NIMService` for NVIDIA's NIM Operator: reach for it where the operator runs, or is declared beside it as `core/ai-model/nvidia-nim-operator`, and affirm that with `infra.control_plane_present`. On a single Docker host, `compose/nvidia-nim` serves it through the NVIDIA Container Toolkit. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint, and where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, while it runs more than one replica on Kubernetes, the Service a scraper discovers every replica behind.

Provenienza

UfficialeFirmato da un’identità di firma che questo catalogo riconosce come editore ufficiale.
Digest del contenuto
sha256:be4033497172ba0ab697392a318f7273bea9b65466407c9f38ae24f5201a3031
Pubblicato
2026-11-10T10:09:26Z
Identità di firma
https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

Firma

  • Ufficialeapplication/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

    sha256:477956835d80e84bd1404eeafc3cab600a40afb0a5ac91daa05c61699609f76a

Riferimento a questa versione

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

Parametri dello sviluppatore

Valori che lo sviluppatore imposta nell’Aggregate.

Estende BaseComponents#/schemas/BaseModelParameter

Dimensionamento dell’operatore

Dimensionamento dell’infrastruttura che l’operatore regola per ogni ambiente.

NomeTipoPredefinitoObbligatorioDescrizione
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).
imagestring—noFull reference of the NIM serving image, including its tag — the image the operator is entitled to pull. NIM images are model-specific, so the reference must be the one for the model named in values.model. There is no default, because access to NIM images is governed by NVIDIA's terms rather than by the catalog, and the catalog therefore names none.
replicasinteger (min 1)1noNumber of serving replicas on Kubernetes, where each replica holds its own model-cache volume and resolves the model profile independently. The NIM Operator realization gives its service one model-cache claim, so it serves one replica and halts the build on more. Docker Compose runs one instance, and the build reports a replica count there as a setting the target omitted.
min_replicasinteger (min 1)—noLower bound for the replica count when horizontal autoscaling is configured. Declared ahead of rendering — each Recipe serves a fixed replica count and none emits an autoscaler yet; 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 — each Recipe serves a fixed replica count and none emits an autoscaler yet; set it to record the intended upper bound.
computeobject—no—
model_cacheobject—noModel-profile cache volume. On Kubernetes it is per replica, sized for the resolved profile, and each serving replica gets its own claim of this size and class. On Docker Compose it is one named volume, which the host creates at its own default, since compose has no volume-size or storage-class concept.
model_artifactstring—noOptional OCI reference to the model weights, an alternative weights source to the model id in values.model. A deterministic pointer the runtime resolves at start; the compiler never fetches it. Declared ahead of rendering — each Recipe resolves the model from values.model and none consumes this pointer yet.
gpuobject—noGPU scheduling for the serving container. NIM is GPU-native, so the serving container always requests at least one GPU; this object tunes the count and selects a GPU runtime class. Requires a GPU-capable cluster and, for runtime_class, a matching RuntimeClass configured in the cluster; on Docker Compose, a host whose NVIDIA GPUs Docker can hand to containers.
image_pull_secretstring—noName of a pre-created Kubernetes image-pull Secret granting access to the NIM container registry. Set this when the NIM image is pulled from a private registry; omit when the cluster can pull it without credentials. On Docker Compose the host pulls with its own registry login, so a Release setting this halts the build there.
source_versionstring—noVersion of NVIDIA's nim-llm chart the `nvidia-nim-chart` realization pulls. Changing it away from that Recipe's default does not guarantee value-mapping compatibility with the new chart. The other realizations pull no chart and ignore it.
control_plane_presentbooleanfalsenoOperator affirmation that NVIDIA's NIM Operator (its controller and its custom-resource definitions) is installed in the target cluster, or is declared beside this component as core/ai-model/nvidia-nim-operator. The NIMService realization halts the build unless this is set true, so a NIMService is never emitted as an orphan into a cluster that cannot reconcile it. The synthesis and chart realizations ignore this flag.
overridesobject (open)—noProvider-specific overrides. Accepts any nested structure. Each Recipe documents which keys it reads.

Contratto di output

Valori che questo componente pubblica verso i consumatori attraverso la topologia uses.

NomeTipoPredefinitoObbligatorioDescrizione
hoststring—sìName the inference server is reached at: the in-cluster DNS name of the Service fronting it on Kubernetes, and the compose service name on Docker Compose. A consumer that composes its own address dials this name; one that does not reads `endpoint` instead.
portinteger—sìTCP port the server accepts model requests on. It follows the port an operator set in the Release, so a consumer reads it rather than assuming a default.
endpointstring—sìFull 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—sìName a request must ask this server for, which is the model profile it loaded rather than a provider catalogue name.
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>/v1/metrics`, on the port the API is served on. Empty while infra.replicas is above one on Kubernetes, when `metrics_service` is published instead; always set on Docker Compose, which runs one instance.
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 while infra.replicas is above one on Kubernetes, beside `metrics_service_port` and `metrics_path`; empty otherwise, and always empty on Docker Compose.
metrics_service_portstring—noName of the port of `metrics_service` that carries the metrics: `http` on the synthesis, `http-openai` on NVIDIA's chart. Set and empty together with `metrics_service`.
metrics_pathstring—noPath the metrics are served at, `/v1/metrics`, published beside `metrics_service`. Empty where the server publishes `metrics_endpoint`, which carries its own path.

Socket uses

I socket di dipendenza attraverso cui questo componente si connette. La destinazione di ogni connessione collegata attraverso un socket deve soddisfare il suo selettore di etichette, e ogni connessione può portare un aggancio negli slot che il socket dichiara. Uno slot obbligatorio deve essere riempito da ogni connessione che l’Aggregate o la Release collega attraverso il socket.

Nessun socket uses dichiarato.

Slot di aggancio

Punti di associazione con nome in cui una Feature si aggancia a questo componente. Il selettore di uno slot limita quale Feature può riempirlo e la sua lista inject indica come può essere consegnato il materiale della Feature. Uno slot obbligatorio deve essere associato su ogni istanza, da un aggancio sull’istanza o a un ambito superiore, e uno slot sensibile è dichiarato per gestire materiale segreto. Una Feature può anche agganciarsi con un nome che nessuno slot dichiara.

SlotSelettoreObbligatorioIniezioneSensibileDescrizione
ngc_api_keyfeature.aggrega.ai/kind=secretnoenvsìOptional NGC API key for authenticating the model-profile download. When attached, the serving container receives it as the NGC_API_KEY environment variable — on Kubernetes read from a Secret by reference, on Docker Compose written into the compose file as delivered, so attach a Compose variable reference such as ${NGC_API_KEY} there to keep the key out of the file. Unattached, the container starts without it. The NIM Operator realization requires it, published under the data key NGC_API_KEY.

Elenco consentito di auto-iniezione

env

Relazioni

Archi tipizzati da e verso altre entità pubblicate. Un riferimento senza una destinazione pubblicata viene mostrato come arco interrotto anziché come link.

Usa

Usato da

Riferimento al concetto Component →

Il catalogo Aggrega — un indice aperto di componenti, funzionalità, ricette e dei plugin che li renderizzano.

Realizzato con il design system condiviso di Aggrega.

Privacy