diff --git a/deployment/vllm/autoscaler.yml b/deployment/vllm/autoscaler.yml index 9c02157..10b2dd4 100644 --- a/deployment/vllm/autoscaler.yml +++ b/deployment/vllm/autoscaler.yml @@ -17,10 +17,15 @@ spec: - name: github-registry-secret containers: - name: autoscaler - image: ghcr.io/wmgeolab/vllm-autoscaler:latest # Update this to your image + image: ghcr.io/wmgeolab/k8s-autoscaler:latest # Update this to your image securityContext: runAsGroup: 50036 runAsUser: 237827 + # runAsNonRoot: true + volumeMounts: + - name: kube-config + mountPath: /app/.kube + readOnly: true resources: requests: memory: "128Mi" # Increased slightly for Python runtime @@ -31,6 +36,8 @@ spec: ports: - containerPort: 8000 env: + - name: KUBECONFIG + value: /app/.kube/config - name: VLLM_SERVICE_HOST value: "vllm-svc" - name: VLLM_SERVICE_PORT @@ -57,6 +64,10 @@ spec: port: 8000 initialDelaySeconds: 5 periodSeconds: 10 + volumes: + - name: kube-config + secret: + secretName: kube-certs --- # Service for the autoscaler apiVersion: v1