# Two-GPU load-balanced scale-out — one vLLM replica per physical GPU, fronted by
# an nginx round-robin LB across the two node IPs.
#
# Why an external nginx LB instead of a ClusterIP Service? Cross-node ClusterIP
# load-balancing DNATs to a backend pod IP on the other node, which rides the pod
# overlay — and that overlay is down under WSL2 mirrored mode (see
# docs/cluster-troubleshooting-log.md). hostNetwork pods bind node IPs on real
# userspace sockets, so an LB that targets node IPs works over plain LAN TCP.
#
# Model is 1.5B at gpu-memory-utilization=0.80 so it fits the SMALLER free pool
# (the desktop 3060 Ti has ~6.7 GB free while driving displays).
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-qwen
  namespace: inference
  labels:
    app: vllm-qwen
spec:
  replicas: 2
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: vllm-qwen
  template:
    metadata:
      labels:
        app: vllm-qwen
    spec:
      runtimeClassName: nvidia
      hostNetwork: true
      dnsPolicy: Default
      # one replica per node: never co-schedule two vLLM pods on the same host
      affinity:
        podAntiAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchLabels:
                  app: vllm-qwen
              topologyKey: kubernetes.io/hostname
      containers:
        - name: vllm
          image: vllm/vllm-openai:latest
          args:
            - --model=Qwen/Qwen2.5-1.5B-Instruct
            - --served-model-name=qwen
            - --gpu-memory-utilization=0.80      # fit the desktop 3060 Ti (~6.7 GB free)
            - --max-model-len=8192
            - --port=8000
          ports:
            - name: http
              containerPort: 8000
          resources:
            limits:
              nvidia.com/gpu: 1
          volumeMounts:
            - name: dshm
              mountPath: /dev/shm
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 40
            periodSeconds: 10
            failureThreshold: 90
      volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 2Gi
---
# nginx round-robin LB across both GPU node IPs. hostNetwork on the desktop so it
# is reachable at 192.168.18.2:8080 and can dial both node IPs over the LAN.
apiVersion: v1
kind: ConfigMap
metadata:
  name: vllm-lb-conf
  namespace: inference
data:
  default.conf: |
    upstream vllm_pool {
        server 192.168.18.2:8000   max_fails=1 fail_timeout=5s;
        server 192.168.18.142:8000 max_fails=1 fail_timeout=5s;
    }
    server {
        listen 8080;
        location / {
            proxy_pass http://vllm_pool;
            proxy_read_timeout 300s;
            proxy_set_header Host $host;
        }
        location = /lb-health { return 200 "ok\n"; }
    }
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-lb
  namespace: inference
  labels:
    app: vllm-lb
spec:
  replicas: 1
  selector:
    matchLabels:
      app: vllm-lb
  template:
    metadata:
      labels:
        app: vllm-lb
    spec:
      hostNetwork: true
      dnsPolicy: Default
      nodeSelector:
        kubernetes.io/hostname: kaiser-desktop     # predictable LB address: 192.168.18.2:8080
      containers:
        - name: nginx
          image: nginx:1.27-alpine
          ports:
            - name: lb
              containerPort: 8080
          volumeMounts:
            - name: conf
              mountPath: /etc/nginx/conf.d
      volumes:
        - name: conf
          configMap:
            name: vllm-lb-conf