CNCF Atlas

English

日本語

English

日本語

Appearance

Getting Started ​

Verified against v0.19.0. Commands assume a local kind cluster and that you serve a predictive model in Standard mode.

Prerequisites ​

  • A Kubernetes cluster. For local testing, kind plus Docker.
  • kubectl configured against that cluster.
  • Outbound network access so the storage-initializer can pull the sample model from public object storage.

Install ​

The exact install asset name changes per release, so follow the official KServe Quickstart Guide for the version you run. The high-level flow is the same across releases.

bash
kind create cluster

The repo ships an install script with explicit mode flags (hack/kserve-install.sh). Standard mode is the lightweight, Knative-free option:

bash
./hack/kserve-install.sh --standard

The script also accepts --knative for serverless mode and --keda for KEDA-based autoscaling. The quickstart guide wraps the same steps in a single curl ... | bash install for your chosen release.

A first working setup ​

Serve the scikit-learn iris model. The manifest below matches the in-repo sample at hack/release/smoke-test-data/sklearn-iris.yaml.

  1. Create the InferenceService.

    yaml
    apiVersion: "serving.kserve.io/v1beta1"
kind: "InferenceService"
metadata:
  name: "sklearn-iris"
spec:
  predictor:
    sklearn:
      storageUri: "gs://kfserving-examples/models/sklearn/1.0/model"

  2. Apply it.

    bash
    kubectl apply -f sklearn-iris.yaml

KServe picks the scikit-learn ServingRuntime by model format, injects a storage-initializer init container that downloads the model from storageUri into /mnt/models, and creates the Deployment, Service, and HPA.

Verify it works ​

Wait for the resource to report Ready, then read its URL:

bash
kubectl get inferenceservice sklearn-iris

The READY column should turn True and the URL column should be populated. The printer columns come from the CRD definition (pkg/apis/serving/v1beta1/inference_service.go:140-147). From there, send a prediction request to that URL following the protocol example in the quickstart guide.

Where to go next ​

  • The KServe Quickstart Guide for the version-pinned install command and a full prediction request.
  • The KServe Admin Guide for production install of Standard, Knative, or ModelMesh modes (linked from the repo README.md).
  • The v0.15 generative AI announcement for the LLMInferenceService path.