# Option A wiring: point the in-cluster `mock-vllm` Service name at a real vLLM
# running on the Windows GPU box over the LAN. Because we reuse the Service name
# and labels, the ServiceMonitor, ScaledObject, dashboard, and load tests all
# target real vLLM unchanged.
#
# A Service with NO selector + a manually-managed Endpoints object lets a
# ClusterIP front an external IP. vLLM serves both the API and /metrics on 8000,
# so Prometheus scrapes the real engine's metrics.
#
# >>> Set the IP below to your Windows box's LAN address. <<<
apiVersion: v1
kind: Service
metadata:
  name: mock-vllm
  namespace: inference
  labels:
    app: mock-vllm        # so the existing ServiceMonitor still selects it
spec:
  ports:
    - name: http
      port: 8000
      targetPort: 8000
---
apiVersion: v1
kind: Endpoints
metadata:
  name: mock-vllm          # must equal the Service name
  namespace: inference
  labels:
    app: mock-vllm
subsets:
  - addresses:
      - ip: 192.168.1.50   # <-- CHANGE ME: Windows box LAN IP (ipconfig)
    ports:
      - name: http
        port: 8000