[Minor Improvements] Vllmruntime Autoscaling in Operator

[aeon-x]

Summary

Three improvements to VLLMRuntime KEDA autoscaling:

1. Fix: resolve --served-model-name override in Prometheus queries — The keepalive trigger queries Prometheus with model="". Previously it always used ModelURL (e.g., meta-llama/Llama-3.1-8B-Instruct), ignoring --served-model-name in ExtraArgs. If a user overrides the served model name, the query returns no data and KEDA may incorrectly scale to zero while traffic is active.
2. Widen keepalive rate window from [1m] to [2m] — A 1-minute window is tight; delayed Prometheus scrapes or brief metric gaps can return 0 and trigger a false scale-to-zero. A 2-minute window is more resilient.
3. Add AvailableReplicas, UpdatedReplicas, UnavailableReplicas to VLLMRuntimeStatus — Exposes deployment replica status directly on the CR, so users can observe scaling progress via kubectl get vr without inspecting the underlying Deployment.

Changes:

• operator/api/v1alpha1/vllmruntime_types.go — Add 3 status fields
• operator/internal/controller/vllmruntime_controller.go — Served-model-name parsing, 2m keepalive window, populate new status fields
• operator/api/v1alpha1/zz_generated.deepcopy.go — Regenerated
• operator/config/crd/bases/production-stack.vllm.ai_vllmruntimes.yaml — Regenerated

Tests Done:

• make generate manifests — regenerated without errors
• make test — all VLLMRuntime and autoscaling tests pass (11/11)
• Deploy with a VLLMRuntime CR using --served-model-name override and verify KEDA ScaledObject uses the correct model name in Prometheus queries
• Verify kubectl get vr shows the new replica status fields

[gemini-code-assist]

Code Review

This pull request adds AvailableReplicas, UpdatedReplicas, and UnavailableReplicas to the VLLMRuntime status and updates the controller to populate these from the Deployment. It also introduces logic to detect --served-model-name overrides in ExtraArgs and increases the Prometheus query interval to 2m. Feedback indicates that the argument parsing logic should be expanded to handle space-separated values to ensure the model name is correctly identified in all common Kubernetes configurations.

[ruizhang0101]

LGTM