derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_conv_add_relu(input_, filter_, bias, stride_w, stride_h,
                             dilation_w, dilation_h, padding, act):
  res = tf.raw_ops.Conv2D(
      input=input_,
      filter=filter_,
      strides=[1, stride_w, stride_h, 1],
      dilations=[1, dilation_w, dilation_h, 1],
      padding=padding)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':

  %input_scale = "tf.Const"() {value = dense<1.0> : tensor<f32>} : () -> tensor<f32>
  %side_input_scale = "tf.Const"() {value = dense<2.0> : tensor<f32>} : () -> tensor<f32>
  %conv2d = "tf._FusedConv2D"(%input, %filter, %bias, %act, %input_scale, %side_input_scale) {

      %outputs_6, %control_7 = tf_executor.island wraps "tf.Const"() {device = "", value = dense<[-1, 16384]> : tensor<2xi32>} : () -> tensor<2xi32>
      %outputs_8, %control_9 = tf_executor.island wraps "tf.Conv2D"(%arg0, %outputs_0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true} : (tensor<16x224x224x3xf32>, tensor<*xf32>) -> tensor<16x112x112x?xf32>

// func.func func_name_${key1}_fn (...) {
//   ...${key2}...
// }
// ```
// The above template with generate two functions by substituting `key1` and
// `key2` with given values.

module {

  for main_op in ["Conv2D", "DepthwiseConv2D", "MatMul"] {
    parameters[
      {"quantized_ops": ["${main_op}", "BiasAdd"], "act_func": "internal_requantize_no_activation_fn", "output_type": "!tf_type.qint8"},

    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  // Canonicalization includes const folding, which is utilized here to optimize
  // away ops that can't get constant folded after PrepareTF pass. For example,
  // tf.Conv2D is split into tf.Transpose and tfl.Conv2D.
  pass_manager->addNestedPass<mlir::func::FuncOp>(
      mlir::createCanonicalizerPass());
  pass_manager->addNestedPass<mlir::func::FuncOp>(mlir::createCSEPass());

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device...

// CHECK:           %[[VAL_9:.*]] = "tf.Transpose"(%[[VAL_7]], %[[VAL_8]]) : (tensor<1x256x256x1xbf16>, tensor<4xi64>) -> tensor<1x1x256x256xbf16>

        ),
    )
    quantization.quantize_saved_model(
        self._input_saved_model_path,
        self._output_saved_model_path,
        config,
    )

    expected_outputs = model.conv2d(input_data)

    root = load.load(self._output_saved_model_path)
    self.assertCountEqual(root.signatures.keys(), {'serving_default'})

    new_outputs = root.signatures['serving_default'](

// NEXT ID: 7
message UnitWiseQuantizationSpec {
  // Quantization unit granularity.
  // NEXT ID: 4
  message QuantizationUnit {
    // Type of the op, ex: Conv2D, MatMul, Einsum... The node_name field can
    // be omitted if it is intended to match all nodes with this type.
    string op_type = 1;
    // Name of the node. This field accepts re2 regex format. If the node name