# KEDA ScaledObject for the real GPU cluster — autoscales the vLLM Deployment on the
# inference signal (queue depth + KV-cache utilization), NOT CPU. Uses KEDA's built-in
# prometheus scaler (two triggers; KEDA scales on whichever is hotter), so it needs no
# custom image. For the single-trigger composite version see the Go external scaler under
# keda-inference-scaler/.
#
# serverAddress is the laptop NODE IP + hostNetwork Prometheus (:9090), not the ClusterIP:
# cross-node ClusterIP routing is down under WSL2 mirrored mode (and NodePort is iptables-
# only, so it isn't delivered either) — so Prometheus is patched to hostNetwork and KEDA
# reaches it over plain LAN TCP (see docs/cluster-troubleshooting-log.md).
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: vllm-qwen
  namespace: inference
spec:
  scaleTargetRef:
    name: vllm-qwen            # the plain vLLM Deployment (one replica per GPU at max)
  minReplicaCount: 1
  maxReplicaCount: 2           # two physical GPUs
  pollingInterval: 15
  cooldownPeriod: 60
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://192.168.18.142:9090
        metricName: vllm_queue_depth
        query: sum(vllm:num_requests_waiting)
        threshold: "3"
        activationThreshold: "1"
    - type: prometheus
      metadata:
        serverAddress: http://192.168.18.142:9090
        metricName: vllm_kv_cache_util
        query: max(vllm:gpu_cache_usage_perc)
        threshold: "0.7"
        activationThreshold: "0.05"