27_wva-variantautoscaling.yaml.j2

{# ============================================================================
   27_wva-variantautoscaling.yaml.j2

   Per-stack VariantAutoscaling resource. Targets the decode Deployment
   created by the llm-d-modelservice chart (name = {model_id_label}-decode
   per chart helper llm-d-modelservice.decodeName + fullnameOverride).

   Labels follow the upstream well-lit-path example in
   llm-d/guides/workload-autoscaling/inference-scheduling-autoscaling/va.yaml
   with values parameterized so each stack emits its own VA regardless of
   how many models share a namespace.

   IMPORTANT: `wva.llmd.ai/controller-instance` label is required when the
   chart is installed with a non-empty wva.controllerInstance. Per the
   v0.6.0 controller's predicate (internal/controller/predicates.go):

     - When CONTROLLER_INSTANCE env is set on the controller, it ONLY
       reconciles VAs whose `wva.llmd.ai/controller-instance` label
       matches that value. Without this label, the predicate returns
       false and the controller silently skips the VA — it'll log
       "No active VariantAutoscalings found" forever.

   The value here MUST equal `wva.controllerInstance` from chart values
   (which we set to wva.namespace | default(namespace.name)) and the
   `controller_instance` label in the HPA selector — all three derive
   from the same expression so they stay aligned.

   The controller (installed once per namespace by step_02) watches this
   namespace and reconciles every VariantAutoscaling resource in it
   that bears the matching controller-instance label.

   Only rendered when wva.enabled is true.
   ============================================================================ #}
{% if wva is defined and wva.enabled | default(false) %}
{% set decode_cfg = decode if decode is defined else {} %}
{% set accel_type = decode_cfg.acceleratorType | default({}) %}
{% set accel_label = (accel_type.labelValues[0] if accel_type.labelValues is defined and accel_type.labelValues else accel_type.labelValue | default('')) %}
{% set accel_name = 'H100' if 'H100' in accel_label
                    else ('A100' if 'A100' in accel_label
                    else ('L40S' if 'L40S' in accel_label
                    else ('MI300X' if 'MI300X' in accel_label
                    else ('G2' if 'G2' in accel_label
                    else '')))) %}
apiVersion: llmd.ai/v1alpha1
kind: VariantAutoscaling
metadata:
  name: {{ model_id_label }}-decode
  namespace: {{ wva.namespace | default(namespace.name, true) }}
  labels:
    inference.optimization/acceleratorName: "{{ accel_name }}"
    llm-d.ai/model: "{{ model.shortName }}"
    llm-d.ai/inference-serving: "true"
    llm-d.ai/guide: "{{ wva.wellLitPath | default('inference-scheduling') }}"
    wva.llmd.ai/controller-instance: "{{ wva.namespace | default(namespace.name, true) }}"
spec:
  scaleTargetRef:
    kind: Deployment
    name: {{ model_id_label }}-decode
  modelID: "{{ model.name }}"
  variantCost: "{{ wva.variantAutoscaling.variantCost | default('10.0') }}"
  minReplicas: {{ wva.variantAutoscaling.minReplicas | default(1) }}
  maxReplicas: {{ wva.variantAutoscaling.maxReplicas | default(10) }}
{% endif %}