apiVersion: v1
kind: ConfigMap
# This ConfigMap defines saturation-based scaling thresholds for model variants.
# Saturation scaling is enabled by default and uses KV cache utilization and queue length
# metrics to determine when replicas are saturated and when to scale up.
#
# Configuration structure:
# - 'default' entry: Global default thresholds applied to all variants
# - Override entries: Per-model/namespace custom thresholds (must include model_id and namespace)
#
# Threshold definitions:
# - kvCacheThreshold: Replica saturated if KV cache utilization >= threshold (0.0-1.0)
# - queueLengthThreshold: Replica saturated if queue length >= threshold (integer)
# - kvSpareTrigger: Scale-up signal if avg spare KV capacity < trigger (0.0-1.0)
# - queueSpareTrigger: Scale-up signal if avg spare queue capacity < trigger (integer)
#
metadata:
  name: wva-saturation-scaling-config
  namespace: workload-variant-autoscaler-system
  labels:
    app.kubernetes.io/name: workload-variant-autoscaler
    app.kubernetes.io/managed-by: kustomize
data:
  # Global defaults applied to all variants unless overridden.
  #
  # The `analyzers` section selects the V2 (token/capacity-based) saturation
  # analyzer — the default since v0.9.0. Remove the `analyzers` section (and the
  # V2-only thresholds) to opt out to the legacy V1 (percentage-based) analyzer.
  # See docs/developer-guide/saturation-scaling-config.md.
  #
  # Each analyzer accepts the ScalingPolicy plugin envelope (proposal #1245):
  #   analyzers:
  #     - type: saturation            # plugin type; `name:` still works as an alias
  #       parameters: { scaleUpThreshold: 0.90 }
  default: |
    analyzers:
      - name: saturation
        score: 1.0
    # V2-only thresholds:
    scaleUpThreshold: 0.85
    scaleDownBoundary: 0.70
    # Shared by V1 and V2:
    kvCacheThreshold: 0.80
    queueLengthThreshold: 5
    # V1-only (ignored by V2):
    kvSpareTrigger: 0.1
    queueSpareTrigger: 3
    # Enable GPU limiter to constrain scaling based on available cluster resources
    # When true, scale-up decisions are limited by available GPU capacity
    enableLimiter: false
    # Enable priority-weighted rescale: under GPU contention the V2 optimizer
    # redistributes the whole budget by priority x demand, reclaiming from
    # lower-priority models so higher-priority work can run. Off by default.
    # Budget-scope flag (read only from this `default` entry): the value here
    # governs the cluster budget; a namespace-local config's `default` governs
    # that namespace's quota budget. Has no effect without a same-scope GPU budget.
    enableRescale: false

  # Per-model overrides inherit the default's analyzer selection (V2) via
  # field-level merge, so they only need the thresholds they change. Under V2,
  # kvCacheThreshold/queueLengthThreshold apply; kvSpareTrigger/queueSpareTrigger
  # are V1-only and are ignored (they take effect only if the default is opted
  # back to V1 by removing its `analyzers:` section).

  # An override entry's DATA KEY must be "{modelID}#{namespace}" — the engine looks
  # up overrides by that key; the inner model_id/namespace fields are informational.

  # Example per-model override for granite model in lab namespace
  # Uncomment and customize as needed
  # "ibm/granite-13b#lab-namespace": |
  #   model_id: ibm/granite-13b
  #   namespace: lab-namespace
  #   kvCacheThreshold: 0.75
  #   queueLengthThreshold: 10

  # Example per-model override for llama model in production namespace
  # Shows partial override (only some fields specified, rest inherit from default)
  # Uses the plugin envelope for the analyzer and an inline scale-to-zero setting
  # (which overrides the wva-model-scale-to-zero-config ConfigMap for this model).
  # "meta/llama-70b#production": |
  #   model_id: meta/llama-70b
  #   namespace: production
  #   priority: 2.0
  #   scaleToZero: { enabled: false }
  #   analyzers:
  #     - type: saturation
  #       parameters: { scaleUpThreshold: 0.90 }

  # Cluster GPU limiter — the sole source that selects the limiter, applied live
  # (no restart) and honored ONLY on the `default` entry above (a cluster-budget
  # scope, like enableRescale). A limiters list selects a SINGLE mode: a quota
  # entry wins over any gpu-inventory entry, so add ONE of the forms below to the
  # `default` entry. With no limiters: declared, the physical-inventory limiter is
  # used.
  #
  #   # (a) physical-capacity limiter:
  #   limiters:
  #     - type: gpu-inventory
  #
  #   # (b) operator-declared quota limiter:
  #   limiters:
  #     - type: quota
  #       name: cluster-h100
  #       scope: cluster
  #       quotas: { H100: 32 }