rewriter.setInsertionPoint(req_op);
    Value new_result = rewriter.create<mlir::stablehlo::UniformDequantizeOp>(
        req_op.getLoc(), func_op.getResultTypes()[0], req_op.getOperand());
    if (func_name.contains("_relu6_")) {
      auto min = rewriter.create<mlir::stablehlo::ConstantOp>(
          req_op.getLoc(), rewriter.getF32FloatAttr(0));
      auto max = rewriter.create<mlir::stablehlo::ConstantOp>(

// CHECK-SAME: attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations"
// CHECK-NEXT: %[[BIASADD_0:.*]] = "tf.BiasAdd"(%[[CONV2D_0]], %arg2)
// CHECK-NEXT: %[[RELU6_0:.*]] = "tf.Relu6"(%[[BIASADD_0]])
// CHECK-NEXT: return %[[RELU6_0]]

// CHECK-LABEL: private @composite_conv2d_with_bias_and_relu_fn_1
// CHECK-NEXT: %[[CONV2D_0:.*]] = "tf.Conv2D"(%arg0, %arg1)

// CHECK: %[[BIASADD_0:.*]] = "tf.BiasAdd"(%[[PARTITIONEDCALL_0]], %[[CONST_0]])
// CHECK: %[[RELU6_0:.*]] = "tf.Relu6"(%[[BIASADD_0]])
// CHECK: %[[PARTITIONEDCALL_1:.*]] = "tf.PartitionedCall"(%arg0, %[[CONST_1]])
// CHECK-SAME: f = @composite_conv2d_fn_1
// CHECK: %[[BIASADD_1:.*]] = "tf.BiasAdd"(%[[PARTITIONEDCALL_1]], %[[CONST_0]])
// CHECK: return %[[RELU6_0]], %[[BIASADD_1]]
// CHECK: }

// CHECK-LABEL: private @composite_conv2d_fn_2

// CHECK-SAME: device = ""
// CHECK:      %[[BIASADD_0:.*]] = "tf.BiasAdd"
// CHECK:      %[[RELU6_0:.*]] = "tf.Relu6"

// CHECK-LABEL: func private @quantized_conv2d_with_bias_and_relu6_fn_0

  elif act == 'RELU6':
    return tf.raw_ops.Relu6(features=res)
  elif act == 'TANH':
    return tf.raw_ops.Tanh(x=res)
  else:
    return res


@tf.RegisterGradient('NewConv2D')
def _conv_add_relu_grad(op: ops.Operation, grad):
  act = op.get_attr('act')
  y = op.outputs[0]
  if act == 'RELU':
    grad = gen_nn_ops.relu_grad(grad, y)
  elif act == 'RELU6':
    grad = gen_nn_ops.relu6_grad(grad, y)

  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %2 = "tfl.add"(%arg0, %arg3) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  // CHECK:          %{{.*}} = "tfl.pseudo_const"() <{value = dense<{{\[\[1, 2\], \[3, 4\], \[5, 6\]\]}}> : tensor<3x2xi32>}>
  // CHECK-NEXT:     [[SUB:%.*]] = tfl.sub %{{.*}}, %{{.*}} {fused_activation_function = "RELU6"} : tensor<3x2xi32>
  // CHECK-NEXT:     [[SCALAR:%.*]] = "tfl.pseudo_const"() <{value = dense<10> : tensor<i32>}> : () -> tensor<i32>

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    func.return %1 : tensor<100xf32>
  }

// CHECK-DAG: %[[ADD:.*]] = stablehlo.add %[[CONV]], %[[BROADCAST]] : tensor<1x3x2x2xf32>
// CHECK-DAG: %[[RELU6:.*]] = stablehlo.clamp %[[CONST_2]], %[[ADD]], %[[CONST_1]] : (tensor<f32>, tensor<1x3x2x2xf32>, tensor<f32>) -> tensor<1x3x2x2xf32>
// CHECK: return %[[RELU6]] : tensor<1x3x2x2xf32>

// -----

func.func @func_conv_batchnorm_relu6_dynamic(%arg_0: tensor<?x3x4x3xf32>) -> (tensor<?x3x2x2xf32>) {

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    func.return %1 : tensor<1xf32>
  }