# Tell Prometheus (Operator) to scrape vLLM's /metrics so KEDA and Grafana
# have the inference SLIs. Requires the Prometheus Operator (PodMonitor CRD).
apiVersion: monitoring.coreos.com/v1
kind: PodMonitor
metadata:
  name: vllm-metrics
  namespace: inference
  labels:
    release: kube-prometheus-stack   # match your Prometheus' podMonitorSelector
spec:
  selector:
    matchLabels:
      serving.kserve.io/inferenceservice: qwen-vllm
  podMetricsEndpoints:
    - port: http                     # the port serving /metrics; adjust to runtime
      path: /metrics
      interval: 15s
# Key series to dashboard in Grafana:
#   vllm:time_to_first_token_seconds   (histogram → TTFT p50/p95)
#   vllm:time_per_output_token_seconds (inter-token latency)
#   vllm:gpu_cache_usage_perc          (KV-cache utilization)
#   vllm:num_requests_waiting          (queue depth)
#   vllm:generation_tokens_total       (throughput → rate())
# SLO suggestion: 95% of requests with TTFT < 1s over a rolling 5m window.