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} {

  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %v = "tf.AddV2"(%y, %z) {__op_key = 3: i32}: (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  return %w, %u, %v : tensor<1xi32>, tensor<1xi32>, tensor<1xi32>
}

// -----

// Test node names are preserved.

// CHECK-LABEL: func @main
func.func @main(%x: tensor<i32>) -> tensor<i32> {
  // CHECK: tf_mlrt.executeop

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.Const"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      func.return %0 : tensor<1xi32>
    }
  })";

absl::StatusOr<std::string> CompileMlirModule(bool compile_to_xla_hlo,
                                              const char* module_str) {

    %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>

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.BadValue"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      func.return %0 : tensor<1xi32>
    }
  })";
  CellReader<int64_t> error(kHasTpuPartitionedCallStreamzName);
  CreateModule(kMlirModuleStr);

  auto result = Run();

// 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>

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.Const"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      func.return %0 : tensor<1xi32>
    }
  })";

  CreateModule(kMlirModuleStr);
  auto result = Run();

  EXPECT_TRUE(result.succeeded());
}

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.Const"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      return %0 : tensor<1xi32>
    }
  })";

  CellReader<int64_t> dynamism_op_counter(kDynamismOpCounterStreamzName);
  CellReader<int64_t> dynamism_function_counter(

  %handle = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "x"} : () -> tensor<!tf_type.resource<tensor<3xf32>>>
  %0 = "tf.ReadVariableOp"(%handle) {_output_shapes = ["tfshape$dim { size: 3 }"], device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<3xf32>>>) -> tensor<3xf32>