spec:
  behavior:
    scaleDown:
      policies:
      - periodSeconds: 3
        type: typeValue
        value: 2
      selectPolicy: selectPolicyValue
      stabilizationWindowSeconds: 3
    scaleUp:
      policies:
      - periodSeconds: 3
        type: typeValue
        value: 2
      selectPolicy: selectPolicyValue
      stabilizationWindowSeconds: 3
  maxReplicas: 3
  metrics:
  - containerResource:

    }

    protected abstract Units<Q> getBaseUnits();

    protected abstract List<Units<Q>> getUnitsForQuantity();

    protected abstract BigDecimal getFactor();

    protected abstract BigDecimal scaleTo(BigDecimal value, Units<Q> units);

    protected String format(BigDecimal value) {
        return value.compareTo(BigDecimal.ONE) == 0 ? displaySingular : displayPlural;
    }

// CHECK-LABEL: @unfuse_batch_norm
// CHECK-SAME: %[[X:[^:[:space:]]+]]
// CHECK-SAME: %[[SCALE:[^:[:space:]]+]]
// CHECK-SAME: %[[OFFSET:[^:[:space:]]+]]
// CHECK-SAME: %[[MEAN:[^:[:space:]]+]]
// CHECK-SAME: %[[VARIANCE:[^:[:space:]]+]]
func.func @unfuse_batch_norm(
    %x: tensor<4x256xf32>, %scale: tensor<256xf32>, %offset: tensor<256xf32>,
    %mean: tensor<256xf32>, %variance: tensor<256xf32>)

            "type": "typeValue",
            "value": "0",
            "averageValue": "0",
            "averageUtilization": 4
          }
        }
      }
    ],
    "behavior": {
      "scaleUp": {
        "stabilizationWindowSeconds": 3,
        "selectPolicy": "selectPolicyValue",
        "policies": [
          {
            "type": "typeValue",
            "value": 2,

          - --container=metrics-server
          - --poll-period=30000
          - --estimator=exponential
          # Specifies the smallest cluster (defined in number of nodes)
          # resources will be scaled to.
          - --minClusterSize={{ metrics_server_min_cluster_size }}
          # Use kube-apiserver metrics to avoid periodically listing nodes.
          - --use-metrics=true
      volumes:

  // quantization dim is 3, so we can cache all the converters.
  ShapedType type = attr.getType();
  std::size_t dim_size = type.getDimSize(quantization_dim_);
  if (dim_size != scales_.size()) {
    return {};
  }
  SmallVector<UniformQuantizedValueConverter, 4> converters;
  converters.reserve(dim_size);
  for (int i = 0, e = dim_size; i != e; ++i) {
    converters.push_back(getPerChunkConverter(i));

// in the range [a, b].
//
// This distribution is *not* normalized (it is not a probability
// distribution). This makes it easier to work with as it's being
// updated.
//
// It is represented as the sum of scaled uniform distribution
// functions and Dirac delta functions (which are treated as
// degenerate uniform distributions).
type mud struct {
	sorted, unsorted []edge

// Converts tf.FusedBatchNormV3 into a sequence of more primitive arithmetic
// operations. Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//
// Let multiplier = scale / sqrt(variance + epsilon),
// to compute
//   (x - mean) * scale / sqrt(variance + epsilon) + offset,
// is then to compute
//   (x * multiplier) + (offset - mean * multiplier).
//

# CHECK-NEXT:          scale: [ 0.015686 ],
# CHECK-NEXT:          zero_point: [ 128 ]
# CHECK-NEXT:        },
# CHECK-NEXT:        has_rank: true
# CHECK-NEXT:      }, {
# CHECK-NEXT:        shape: [ 4 ],
# CHECK-NEXT:        type: UINT8,
# CHECK-NEXT:        buffer: 2,
# CHECK-NEXT:        name: "input1",
# CHECK-NEXT:        quantization: {
# CHECK-NEXT:          scale: [ 0.023529 ],

    prometheus.io/scrape: "true"

  # Additional labels to apply on the pod level
  podLabels: {}

  # Pod resource configuration
  resources:
    requests:
      cpu: 200m
      # Ztunnel memory scales with the size of the cluster and traffic load
      # While there are many factors, this is enough for ~200k pod cluster or 100k concurrently open connections.
      memory: 512Mi