// point constant.
def : Pat<(TFL_DequantizeOp
             (TFL_QuantizeOp (Arith_ConstantOp F32ElementsAttr:$cst), $qt)),
          (TFL_ConstOp $cst)>;

// Transpose conv supports hybrid computation with quantized weights.
def FoldQuantWeightsIntoTposeConv : Pat<
  (TFL_TransposeConvOp
    $output_shape,
    (TFL_DequantizeOp $quant_weights),
    $quant_input,
    $bias, $padding, $stride_h, $stride_w, $faf),

		return scaleFactor
	}
	weight := uint32(math.MaxUint32)
	if ep.GetLoadBalancingWeight().Value < math.MaxUint32/scaleFactor {
		weight = ep.GetLoadBalancingWeight().Value * scaleFactor
	}
	return weight
}

// Apply the weight for this endpoint to the network gateways.

	}
	return postProcessNamedEdgeMap(weight, totalWeight)
}

func sortByWeight(edges []NamedCallEdge, weight map[NamedCallEdge]int64) {
	sort.Slice(edges, func(i, j int) bool {
		ei, ej := edges[i], edges[j]
		if wi, wj := weight[ei], weight[ej]; wi != wj {
			return wi > wj // want larger weight first
		}
		// same weight, order by name/line number
		if ei.CallerName != ej.CallerName {

    // determined. The activation and weight are quantized to INT8 while bias is
    // quantized to INT32.
    METHOD_STATIC_RANGE_INT8 = 2;

    // Dynamic range quantization. Quantized tensor values' ranges are
    // determined in the graph executions. The weights are quantized during
    // conversion.
    METHOD_DYNAMIC_RANGE_INT8 = 3;

    // Weight-only quantization. Only weights are quantized during conversion.

    $constant_operand_indices are the indices of the inputs that are constant to the TPU program (e.g. weights in inference), the rest of the inputs are input tensors.
    constant_operand_indices is sorted in ascending order.
    $operands_with_static_shape are indices of operands that are tagged with a maximum static shape.

      enable_per_channel_quantized_weight_;
  // Change this to user-given bit width once we have custom configuration.
  options.bit_width_ = 8;

  // Insert quantization parameters for weights for ops with `weight_only_ptq`
  // attribute.
  pm.addNestedPass<func::FuncOp>(createInsertWeightParamPass());

  // PrepareQuantizePass uses SymbolTable to fetch relevant GEMM ops for

                     "Uses TF Uniform Quantized ops"))};

  Option<int64_t> min_num_elements_for_weights_{
      *this, "min-num-elements-for-weights", llvm::cl::init(0),
      llvm::cl::desc("The minimum required number of elements in a weight "
                     "array to apply quantization.")};

  Option<QuantMethod> quantization_method_{
      *this, "quantization-method",

    $constant_operand_indices are the indices of the inputs that are constant to the TPU program (e.g. weights in inference), the rest of the inputs are input tensors.
    constant_operand_indices is sorted in ascending order.
    $operands_with_static_shape are indices of operands that are tagged with a maximum static shape.

	// Type selects which strategy to run.
	Type ScoringStrategyType

	// Resources to consider when scoring.
	// The default resource set includes "cpu" and "memory" with an equal weight.
	// Allowed weights go from 1 to 100.
	// Weight defaults to 1 if not specified or explicitly set to 0.
	Resources []ResourceSpec

	// Arguments specific to RequestedToCapacityRatio strategy.

			continue
		}

		entry, weight := split(entry, ';')

		// Scan the language.
		t, err := Parse(entry)
		if err != nil {
			id, ok := acceptFallback[entry]
			if !ok {
				return nil, nil, err
			}
			t = makeTag(language.Tag{LangID: id})
		}

		// Scan the optional weight.
		w := 1.0
		if weight != "" {
			weight = consume(weight, 'q')
			weight = consume(weight, '=')