openova/platform/kserve

KServe

Kubernetes-native model serving. Application Blueprint (see docs/PLATFORM-TECH-STACK.md §4.6). Used by bp-cortex to serve LLMs via vLLM, embedding models via BGE, and any custom inference workload.

Status: Accepted | Updated: 2026-04-30

---

Blueprint chart

This folder ships an umbrella Helm chart at chart/ that wraps the upstream kserve/kserve chart (v0.16.0 — latest version published on the official OCI registry as of 2026-04-30) under dependencies:. Catalyst-curated overlay templates render alongside:

• chart/templates/networkpolicy.yaml — locks the controller-manager namespace down (DEFAULT FALSE).
• chart/templates/servicemonitor.yaml — controller-manager metrics scrape (DEFAULT FALSE per docs/BLUEPRINT-AUTHORING.md §11.2; Capabilities-gated).
• chart/templates/hpa.yaml — controller-manager Deployment HPA (DEFAULT FALSE; controller is leader-elected).

Catalyst defaults:

• kserve.controller.deploymentMode: RawDeployment — KServe writes plain Deployment+Service+HPA per InferenceService (no Knative hop on the hot path).
• kserve.controller.gateway.ingressGateway.enableGatewayApi: true + className: cilium — Catalyst's istio-less Cilium native Gateway-API path.
• kserve.controller.gateway.disableIstioVirtualHost: true — Knative-Istio is NOT installed.
• bp-knative is still installed (declared as a hard dependency in blueprint.yaml) so per-InferenceService annotation serving.kserve.io/deploymentMode: Serverless opts in to scale-to-zero on a per-tenant basis without infra changes.

---

Overview

KServe provides standardized model serving on Kubernetes with support for multiple ML frameworks, autoscaling, and inference graphs.

flowchart TB
    subgraph KServe["KServe"]
        Controller[KServe Controller]
        Predictor[Predictor]
        Transformer[Transformer]
        Explainer[Explainer]
    end

    subgraph Runtimes["Serving Runtimes"]
        vLLM[vLLM]
        TorchServe[TorchServe]
        Triton[Triton]
        SKLearn[SKLearn]
    end

    subgraph Knative["Knative Serving"]
        Autoscale[Autoscaling]
        Revisions[Revisions]
    end

    Controller --> Predictor
    Controller --> Transformer
    Controller --> Explainer
    Predictor --> Runtimes
    Runtimes --> Knative

---

Why KServe?

Feature	Benefit
Multi-framework	TensorFlow, PyTorch, ONNX, vLLM, etc.
Autoscaling	Scale-to-zero via Knative
InferenceService	Standardized deployment pattern
Inference Graph	Multi-model pipelines
Model explainability	Integrated explainers

---

Components

Component	Purpose
InferenceService	Model deployment abstraction
ServingRuntime	Framework-specific runtime
InferenceGraph	Multi-model orchestration
ClusterStorageContainer	Model storage configuration

---

Serving Runtimes

Runtime	Use Case
vLLM	LLM inference (recommended)
TorchServe	PyTorch models
Triton	Multi-framework, high performance
SKLearn	Scikit-learn models
XGBoost	Gradient boosting models
ONNX	ONNX format models

---

Configuration

InferenceService Example

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: llm-service
  namespace: ai-hub
spec:
  predictor:
    model:
      modelFormat:
        name: vllm
      runtime: vllm-runtime
      storageUri: pvc://model-cache/models/qwen-32b
      resources:
        requests:
          cpu: "4"
          memory: 32Gi
          nvidia.com/gpu: "2"
        limits:
          cpu: "8"
          memory: 64Gi
          nvidia.com/gpu: "2"

ServingRuntime for vLLM

apiVersion: serving.kserve.io/v1alpha1
kind: ServingRuntime
metadata:
  name: vllm-runtime
spec:
  supportedModelFormats:
    - name: vllm
      autoSelect: true
  containers:
    - name: kserve-container
      image: vllm/vllm-openai:latest
      args:
        - --model=$(MODEL_ID)
        - --tensor-parallel-size=2
        - --max-model-len=32768
      resources:
        requests:
          nvidia.com/gpu: "2"

---

Inference Graph

Multi-model pipeline for complex inference:

apiVersion: serving.kserve.io/v1alpha1
kind: InferenceGraph
metadata:
  name: rag-pipeline
spec:
  nodes:
    root:
      routerType: Sequence
      steps:
        - serviceName: embedder
        - serviceName: retriever
        - serviceName: llm
    embedder:
      serviceName: bge-embedder
    retriever:
      serviceName: vector-search
    llm:
      serviceName: qwen-llm

---

GPU Scheduling

# Node selector for GPU nodes
spec:
  predictor:
    nodeSelector:
      nvidia.com/gpu.product: NVIDIA-A10
    tolerations:
      - key: nvidia.com/gpu
        operator: Exists
        effect: NoSchedule

---

Model Storage

PVC-based Storage

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: model-cache
  namespace: ai-hub
spec:
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 100Gi
  storageClassName: oci-bv

S3-based Storage (SeaweedFS)

apiVersion: serving.kserve.io/v1alpha1
kind: ClusterStorageContainer
metadata:
  name: seaweedfs-storage
spec:
  container:
    name: storage-initializer
    image: kserve/storage-initializer:latest
    env:
      - name: AWS_ACCESS_KEY_ID
        valueFrom:
          secretKeyRef:
            name: seaweedfs-credentials
            key: accesskey
      - name: AWS_SECRET_ACCESS_KEY
        valueFrom:
          secretKeyRef:
            name: seaweedfs-credentials
            key: secretkey
      - name: S3_ENDPOINT
        value: http://seaweedfs.storage.svc:8333

---

Monitoring

Metric	Query
Inference latency	kserve_inference_duration_seconds
Request count	kserve_inference_count
GPU utilization	DCGM_FI_DEV_GPU_UTIL
Model load time	kserve_model_load_duration_seconds

---

Consequences

Positive:

• Standardized model deployment
• Multi-framework support
• Autoscaling via Knative
• Inference graphs for pipelines
• GPU scheduling support

Negative:

• Complexity for simple deployments
• Requires Knative
• Learning curve for KServe concepts

---

Part of OpenOva