Pular para o conteúdo principal

Inference routing

Configure Kubernetes Gateway API or standalone inference routing for AI workloads.

Agentgateway supports the Kubernetes Gateway API Inference Extension in two deployment modes.

Kubernetes Gateway API mode​

In Kubernetes Gateway API mode, agentgateway runs as the gateway data plane for Gateway API resources. You install the Inference Extension CRDs, create an InferencePool, and route to that pool from an HTTPRoute. The Endpoint Picker Extension (EPP) acts as an extension service that selects the best model server endpoint for each inference request.

Use this mode when you want Gateway API integration, InferencePool resources, traffic splitting, route matching, and other Kubernetes networking features.

Set up Kubernetes inference routing

Standalone request scheduler mode​

In standalone request scheduler mode, agentgateway runs as a sidecar proxy with the EPP. The proxy and EPP communicate over localhost, and agentgateway uses its standalone inferenceRouting local configuration to route requests to a synthetic service before consulting the EPP for endpoint selection.

Use this mode for single-tenant or job-scoped workloads where deploying a full Gateway API stack would add unnecessary operational overhead. In this mode, the upstream standalone Helm chart can deploy agentgateway as the sidecar proxy with proxyType: agentgateway.

Standalone request scheduler mode does not support InferencePool. The standalone configuration must define a top-level synthetic service, such as a services entry, and the route backend must reference that service. When EPP owns endpoint discovery, set destinationMode: passthrough so EPP-selected destinations can be forwarded to directly without matching local workload endpoint data.

For example, the standalone agentgateway configuration defines the synthetic service in services, and the route backend references it as default/my-model.

config.yaml

services:
- name: my-model
  namespace: default
  hostname: my-model
  vips: []
  ports:
    8000: 8000

binds:
- port: 8081
  listeners:
  - name: default
    protocol: HTTP
    routes:
    - name: standalone-epp
      matches:
      - path:
          pathPrefix: /v1/chat/completions
      backends:
      - service:
          name: default/my-model
          port: 8000
        policies:
          inferenceRouting:
            endpointPicker:
              host: 127.0.0.1:9002
            destinationMode: passthrough