%4 = "tf.BiasAdd"(%3, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %5 = "tf.Relu6"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

    func.return %2, %5 : tensor<*xf32>, tensor<*xf32>
  }

  func.func @float_conv_1(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

      %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
      %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>

    // expected-remark@+1 {{Cost: 1}}
    %14 = "tf.Reshape"(%7, %13) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<?x8x?x?xf32>, tensor<4xi32>) -> tensor<?x8x?x?xf32>
    // expected-remark@+1 {{Cost: 8}}
    func.return %11, %14 : tensor<4xi32>, tensor<?x8x?x?xf32>
}

// CHECK-LABEL: test_expensive_ops

          %8 = "tf.C"(%arg1) {is_stateless = true} : (tensor<i32>) -> tensor<i32>
          %9 = "tf.D"(%arg1, %arg2) {_xla_outside_compilation = "cluster1", is_stateless = true} : (tensor<i32>, tensor<3xf32>) -> tensor<3xf32>
          "tf.Yield"(%8, %9) : (tensor<i32>, tensor<3xf32>) -> ()
        }) {is_stateless = true} : (tensor<i32>, tensor<3xf32>) -> (tensor<i32>, tensor<3xf32>)

      } : (tensor<?x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<?xf32>
      func.return %1 : tensor<?xf32>
    }
}

// -----

// expected-error@+1 {{Input shapes provided but no `main` function found.}}
module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @non_main(%arg0 : tensor<?xf32>) -> tensor<?xf32> {

module {
  func.func @wrap_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    %0 = "tf.PartitionedCall"(%arg0, %arg1) <{f = @composite_conv2d_with_relu6_fn}> {_tfl_quant_trait = "fully_quantizable"}
          : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  func.func @no_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {

      return %2 : tensor<?x2xf32>
    }
    func.func private @composite_dot_general_fn_1(%arg0: tensor<?x2xf32>, %arg1: tensor<2x2xf32>, %arg2: tensor<2xf32>) -> tensor<?x2xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {

    %r_max_without_zero = "tf.Max"(%input_1d, %dim) { keep_dims = true }: (tensor<?xf32>, tensor<1xi64>) -> tensor<1xf32>
    %r_max = "tf.Maximum"(%zero, %r_max_without_zero) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %r_min_without_zero = "tf.Min"(%input_1d, %dim) { keep_dims = true }: (tensor<?xf32>, tensor<1xi64>) -> tensor<1xf32>

// CHECK-LABEL: @main
// CHECK: %[[RES0:.*]] = "quantfork.stats"(%arg0) <{layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>}> : (tensor<1x512x672x8xf32>) -> tensor<1x512x672x8xf32>

  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x512x672x8xf32>) -> tensor<1x512x672x8xf32>
  func.return %0 : tensor<1x512x672x8xf32>

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<3xf32>) -> tensor<*xf32>
    %2 = "tf.PartitionedCall"(%arg0, %cst_2) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_1} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
    func.return %1, %2: tensor<*xf32>, tensor<*xf32>